Complete Hard Disk Encryption

Complete Hard Disk Encryption Using FreeBSD's
GEOM Framework
Marc Schiesser
m.schiesser [at] quantentunnel.de
October 20th 2005
Abstract
Most technologies and techniques intended for securing digital
data focus on protection while the machine is turned on – mostly by
defending against remote attacks. An attacker with physical access
to the machine, however, can easily circumvent these defenses by
reading out the contents of the storage medium on a different, fully
accessible system or even compromise program code in order to
leak encrypted information.
Especially for mobile users, that threat is real. And for those
carrying around sensitive data, the risk is most likely high.
This paper describes a method of mitigating that particular risk
by protecting not only the data through encryption, but also the
applications and the operating system from being compromised
while the machine is turned off.
The platform of choice will be FreeBSD, as its GEOM framework
provides the flexibility to accomplish this task. The solution does
not involve programming, but merely relies on the tools already
provided by FreeBSD.
Table of Contents
1 Background & motivation......................................................................................................2
2 Partial disk encryption...........................................................................................................3
2.1 Filebased encryption.....................................................................................................4
2.2 Partitionbased encryption............................................................................................5
2.3 The leakage risk...............................................................................................................5
2.4 New attack vectors..........................................................................................................6
3 Complete disk encryption.....................................................................................................6
3.1 Tools provided by FreeBSD............................................................................................6
3.2 The problem with complete disk encryption...............................................................7
3.3 Requirements..................................................................................................................8
3.4 Complete hard disk encryption using GBDE................................................................8
3.4.1 Erasing previously stored data...............................................................................8
3.4.2 Initialization & the lockfile......................................................................................9
3.4.3 Attaching the encrypted medium..........................................................................9
3.4.4 Partitioning............................................................................................................10
3.4.5 Creating the filesystem..........................................................................................11
3.4.6 Installing FreeBSD.................................................................................................11
3.4.7 Preparing the removable medium.......................................................................12
3.4.8 The kernel modules...............................................................................................12
3.4.9 The problem with GBDE.......................................................................................13
3.4.10 The memory disk.................................................................................................13
3.4.11 Populating the memory disk filesystem............................................................14
3.4.12 The booting process............................................................................................14
3.4.13 Creating the symlinks..........................................................................................15
3.4.14 Integrating the memory disk image...................................................................15
3.4.15 The swap partition...............................................................................................16
3.4.16 Postinstallation issues.......................................................................................16
3.5 Complete hard disk encryption using GELI................................................................16
3.5.1 Readying the hard disk..........................................................................................17
3.5.2 Improvements and new problems with GELI.....................................................17
3.5.3 Initializing, attaching and partitioning................................................................18
3.5.4 Filesystem creation and system installation.......................................................19
3.5.5 The removable medium........................................................................................19
3.5.6 Mounting the encrypted partition.......................................................................19
4 Complete hard disk encryption in context.........................................................................20
4.1 New defenses & new attack vectors – again................................................................20
4.2 Tradeoffs......................................................................................................................22
4.3 GBDE vs. GELI...............................................................................................................23
5 Conclusion............................................................................................................................23
References & further reading.................................................................................................24
1 Background & motivation
As more and more data enters the digital world, appropriate measures must be taken in
order to protect it.
Considering the everincreasing number of networked devices and the inherent
exponential growth of the Internet, it is imperative that a large amount of effort go into
securing devices against remote attacks. Common technologies and techniques include
firewalls, intrusion detection systems (IDS), encryption of all kinds of network
transmissions as well as hardening network stacks and fixing buffer overflows.
At the same time, we are witnessing increasingly sophisticated and complex mobile
devices such as PDAs, smartphones and cell phones becoming pervasive and assuming
all kinds of important tasks. Between the generalpurpose laptop and the (once) special
purpose cell phone, pretty much anything in between is available.
As people use these devices, they also generate data – either explicitly or implicitly.
Explicitly stored data might for example include: entering a meeting into the electronic
schedule, storing a telephone number and associating a name with it, or saving an email
message draft in order to finish it later.
But then there is also the data which is stored implicitly. Examples include the
history of the telephone numbers called or received, browser caches, recently accessed
files, silently by the software archived or backedup data such as email messages, log
files and so on.
Even if the user remembers to delete the explicitly stored files after they are no longer
needed, it is possible to trace a lot of his or her activity on the device by looking at the
aforementioned, implicitly stored data. The more sophisticated the device is, the more
such data will usually be generated, mostly without the user's knowledge.
In terms of performance, laptop computers hardly lag behind their desktop
counterparts – enabling them to run the same powerful and complex software. It also
means that the users tend to generate far more data – both explicitly and implicitly –
than on simpler devices.
In addition to being exposed to remote attacks, laptop users are also faced with an
increased exposure of the machine itself. While stationary computers are physically
accessible by usually only a limited number of people, a laptop computer is intended to
be used anywhere and anytime.
This paper does not try to provide any solutions to mitigating the risks of remote
attacks. Instead, it concentrates on the risks posed by attackers with physical access to
the device. An attacker with physical access to a machine can either:
• boot his own operating system, thus circumventing restrictions such as login
procedures, filesystem and network access control and sandboxes
• or remove the hard disk from the targeted machine and install it in a system
which is under the control of the attacker – in case the target's booting
sequence is protected (e.g. by a BIOS password)
Unfortunately, most people and companies take quite lax an approach when it comes to
protecting their data instorage, while the machine is turned off. The following quotes
illustrate just how serious a problem the lack of instorage encryption can become:
• „Thieves stole computer equipment from Fort Carson containing soldiers'
Social Security numbers and other personal records, the Army said ...” [Sarche,
2005]
• „Personal devices "are carrying incredibly sensitive information," said Joel
2
Yarmon, who, as technology director for the staff of Sen. Ted Stevens (R
Alaska), had to scramble over a weekend last month after a colleague lost one of
the office's wireless messaging devices. In this case, the data included "personal
phone numbers of leaders of Congress. . . . If that were to leak, that would be
very embarrassing," Yarmon said.” [Noguchi, 2005]
• „A customer database and the current access codes to the supposedly secure
Intranet of one of Europe's largest financial services group was left on a hard
disk offered for sale on eBay.” [Leyden, 2004]
• „ ... Citigroup said computer tapes containing account data on 3.9 million
customers, including Social Security numbers, were lost by United Parcel
Service.” [Reuters, 2005]
• „Earlier this year, a laptop computer containing the names and Social Security
numbers of 16,500 current and former MCI Inc. employees was stolen from the
car of an MCI financial analyst in Colorado. In another case, a former Morgan
Stanley employee sold a used BlackBerry on the online auction site eBay with
confidential information still stored on the device. And in yet another incident,
personal information for 665 families in Japan was recently stolen along with a
handheld device belonging to a Japanese powercompany employee.”
[Noguchi, 2005]
• „ ... trading firm Ameritrade acknowledged that the company that handles its
backup data had lost a tape containing information on about 200,000
customers. ” [Lemos, 2005]
• „MCI last month lost a laptop that stores Social Security numbers of 16,500
current and former employees. Iron Mountain, an outside data manager for
Time Warner, also lost tapes holding information on 600,000 current and
former Time Warner workers.” [Reuters, 2005]
Even though the number of press articles reporting damage due to stolen mobile
computers – or more specifically: storage media – does not reach the amount of publicity
that remotely attacked and compromised machines provoke, it must also be taken into
account that data on a laptop does not face as much exposure as it does on an Internet
server.
A laptop computer can be insured and data regularly be backed up in order to limit
the damage in case of loss or theft; but protecting the data from unauthorized access
requires a different approach.
2 Partial disk encryption

Encryption of instorage data (as opposed to intransmission) is not a completely new
idea, though. Several tools for encrypting individual files have been around for quite
some time. Examples include the famous PGP (Pretty Good Privacy) as well as its free
counterpart GnuPG and the somewhat less known tools AESCrypt1 and NCrypt2.
More sophisticated approaches aim towards encrypting entire partitions. The
1 http://aescrypt.sourceforge.net/
2 http://ncrypt.sourceforge.net/
3
vncrypt3 project is an example that takes this approach.
2.1 File-based encryption

The idea is that the user can decide for each file individually, whether and how it is to be
encrypted. This has the following implications:
• CPU cycles can be saved on data that the user decides is not worth the effort.
This is an advantage, since encryption requires a lot of processing power. It also
allows the user to choose different keys for different files (although reality
usually reflects the opposite phenomenon).
• Meta data is not encrypted. Even if the file's contents are sufficiently protected,
information such as file name, ownership, creation and modification date,
permissions and size are still stored in the clear. This represents a risk which is
not to be underestimated.
The usability of instorage encryption largely depends on how transparent the
encryption and decryption process is performed to the user. In order to minimize user
interaction, the relevant system calls must be modified to perform the cryptographic
functions accordingly. That way, neither the user nor the applications must make any
additional effort to process encrypted file content, since the kernel will take care of this
task.
If system call modification is not going to take place, any program required to
process encrypted data must either be modified to perform the necessary cryptographic
functions itself or it must rely on an external program for that task. This conversion
between cipher text and plain text – and vice versa – is hardly possible without requiring
any user interaction.
Scenario: filebased encryption of huge files
The file might be a database or multimedia container: if the cryptographic functions
are not performed onthefly (usually through modified system calls), the entire file
content must be temporarily stored in plain text – therefore consuming twice the space.
Then the unencrypted copy must be opened by the application. After the file has
been closed, it obviously must be encrypted again – unless no modification has taken
place. The application will therefore save the data in plain text, which must then be
encrypted and written out in its cipher text form again – but by a program capable of
doing the appropriate encryption.
The unencrypted (temporary) copy could of course just be unlinked (removed from
the name space), but in that case the unencrypted data would still remain on the
medium until physically completely overwritten. So, if one wants to really destroy the
temporary copy, several overwrites are required – which can consume a lot of time with
large files. Therefore, a lot of unnecessary I/O must be performed.
Scenario: filebased encryption of many files
If one wants to encrypt more than just a small bunch of files, it actually does not
matter how small or large they are – the procedure described above still must be adhered
to unless encryption and decryption is performed onthefly.
Aside from its lack of scalability, filebased encryption also suffers from the leakage
3 http://sourceforge.net/projects/vncrypt/
4
risk “phenomenon” – which will be discussed in chapter 2.3. In most cases, encryption is
therefore either abandoned or the following, more effective and efficient scheme is
chosen.
2.2 Partition-based encryption

Obviously, creating a temporary plain text copy of an entire partition each time the data
is accessed, is hardly a sane solution. The system must therefore be able to perform the
encryption and decryption onthefly, as it has been implemented in the FreeBSD kernel
for GBDE [Kamp, 2003a] and GELI [Dawidek, 2005a] and cgd(4) in NetBSD
[Dowdeswell & Ioannidis, 2003]. A few third party addons also exist. One example of this
is the aforementioned vncrypt, which was developed at Sourceforge.
vncrypt is, however, in a further sense still filebased, because the encrypted
partition is only a mounted pseudodevice created via the vn(4) facility from a regular
file. This file holds all the partition's data in encrypted form – including meta data.
OpenBSD's vnconfig(8) provides a similar feature [OpenBSD, 1993].
One aspect associated with partitionbased encryption is that its setup process is
usually more extensive than it is for filebased encryption. But once it has been done,
partitionbased encryption is far superior to the filebased encryption scheme. All data
going to the particular partition is – by default – stored encrypted. As both encryption
and decryption is performed transparently to the user and onthefly, it is also feasible to
encrypt both large amounts of files and large amounts of data.
But unfortunately, this scheme it not perfect either.
2.3 The leakage risk

As obvious as it may sound, partitionbased encryption protects only what goes onto the
encrypted partition. The following scenario highlights the particular problem.
Scenario: editing a sensitive document stored on an encrypted partition
A mobile user needs to have a lot of data at his immediate disposal. Since some
information is sensitive, he decides to put it on an encrypted partition.
Unfortunately, the encryption becomes basically useless as soon as encrypted files
are opened with applications that create temporary copies of the files currently being
worked on, often in the /tmp directory. So unless the user happens to have /tmp
encrypted, his sensitive data is leaked to an unencrypted part of the medium. Even if the
application deletes the temporary copy afterwards, the data still remains on the medium
until it is physically overwritten. Meta data such as file name, size and ownership may
also have leaked and may therefore remain accessible for some time.
This phenomenon happens equally implicitly with printing. Even if the application
itself does not leak any data, the spooler will usually create a Postscript document in a
subdirectory of /var/spool/lpd/, which is not encrypted unless specifically done so.
Even though it is possible to symlink the “hot” directories such as /tmp, /var/tmp, as
well as the complete /home or /var/spool/lpd/ to directories on the encrypted partition,
the leakage risk can never be avoided completely. It is something that users of partition
based encryption just have to be aware of and learn to live with by minimizing the
amount of leaked data as much as possible.
5
The leakage risk is also another reason why filebased encryption is virtually useless.
While this issue is certainly a problem for sensitive data, there is a far bigger problem,
which so far has been quietly ignored.
2.4 New attack vectors

The point of storing data is to be able to retrieve it at some later date. So far, everything
that was discussed, was based on the assumption that both the OS and the applications
were stored unencrypted – there is also no point in doing otherwise as long as the data
itself is not encrypted:
• if data cannot4 be destroyed, stolen or modified remotely, a dedicated attacker
will find a way to gain local (physical) access to the system
• if login procedures, filesystem access control and other restrictions imposed by
the OS and applications cannot be defeated or circumvented, the attacker will
boot his/her own OS
• if the booting sequence on the machine is protected, the attacker will remove
the hard disk and access it from a system under his control
• if the data on the hard disk is encrypted and a bruteforce attack is not feasible,
then the attacker will most likely5 target the OS and/or the applications
The key motivation behind complete disk encryption is illustrated in the last point: the
OS and the applications are now the target. Instead of breaking the encryption, an
attacker can try and subvert the kernel or the applications, so they leak the desired data
or the encryption key.
The goal is therefore to encrypt the OS and all the applications as well. Just as any
security measure that is taken, this scheme involves tradeoffs, such as less convenience
and decreased performance. These issues will be discussed later. Every user considering
this scheme must therefore decide for him or herself, whether the increase in security is
worth the tradeoffs.
3 Complete disk encryption
3.1 Tools provided by FreeBSD

The platform of choice here is FreeBSD, because it comes with a modular, very powerful
I/O framework called GEOM [Kamp, 2003b] since the release of the 5.x branch. The 5.x
branch underwent several major changes compared to the 4.x branch and was not
declared STABLE until the 5.3RELEASE in November 2004. The 5.x branch did,
however, feature a GEOM class and a corresponding userland utility called GBDE
(GEOM Based Disk Encryption) as early as January 2003 when 5.0RELEASE came out.
GBDE was specifically designed to operate on the sector level and is therefore able to
4 perfect security is not possible; therefore 'cannot' should rather be read as 'cannot easily enough'
5 tampering with the hardware is of course also possible, for example with a hardware keylogger;
defending against this kind of attack is not discussed in this paper
6
encrypt entire partitions and even hard disks or other media.
When the 5.x branch was finally declared STABLE and therefore ready for
production use, 6.x became the new developer branch, carrying all the new, more
disruptive features. Into this branch added was also a new module and userland utility
called GELI [Dawidek, 2005b]. In addition to containing most of the GBDE features, GELI
was designed to enable the kernel to mount the root filesystem (/) from an encrypted
partition. GBDE does not allow to do this and therefore requires a “detour“ in order to
make complete hard disk encryption work.
This paper will discuss the realization of complete hard disk encryption with both
tools without having to rely on programming. GELI is a more elegant solution, because it
was designed with this application in mind. GBDE, on the other hand, has seen more
exposure because it has been available for much longer then GELI and therefore is more
likely to have received more testing. Using GBDE for complete hard disk encryption also
illustrates some interesting problems inherent with the booting process and how these
can be solved.
Which approach is in the end chosen, is left to the user. The following table lists the
most important features of GBDE and GELI [Dawidek, 2005b].
GBDE GELI
First released in FreeBSD 5.0 6.0
Cryptographic algorithms AES AES, Blowfish, 3DES
Variable key length No Yes
Allows kernel to mount encrypted root partition No Yes
Dedicated hardware encryption acceleration No Yes, crypto(9)
Passphrase easily changeable Yes Yes
Filesystem independent Yes Yes
Automatic detach on last close No Yes
Table 1: the most important GBDE and GELI features
3.2 The problem with complete disk encryption

There are cases in which it is desirable to encrypt the whole hard disk – especially with
mobile devices. This also includes the encryption of the kernel and the boot loader.
Today's computers, however, cannot boot encrypted code. But if the boot code is not
encrypted, it can easily be compromised. The solution is therefore to store all code
necessary for booting and then mounting the encrypted hard disk partition on a medium
that can be carried around at all times.
While virtually any removable medium is easier to carry around than a fixed one, USB
memory sticks are currently the best solution. They provide plenty of space at affordable
prices, are easily rewritable many times and easy to use since operating systems treat
them like a hard disk. But most importantly, they are small and light.
Obviously, putting the boot code on a removable medium instead of the fixed hard
disk does not solve the problem of compromise – the risk is simply shifted toward the
removable medium. But since that medium can be looked after a lot more easily, there is
a considerable benefit to the user.
7
3.3 Requirements
Independent of whether GBDE or GELI is used, the following things are required:
• A bootable, removable medium. It will carry the boot code as well as the kernel.
This medium is preferably a USB memory stick, because it is small, light and
offers a lot of easily rewritable space.
• The device intended for complete disk encryption. Is is very important that this
device is capable of booting from the removable medium mentioned above.
Especially older BIOSes may not be able to boot from USB mass storage.
Bootable CDs will probably work on most machines. Although they work
equally well (r/w access is not a requirement for operation), they are harder to
set up and maintain.
• In order to set up and install everything, a basic FreeBSD system is required.
The FreeBSD installation discs carry a “live filesystem” – a FreeBSD system
which can be booted directly from the CD. It can be accessed via the
sysinstall menu entry 'Fixit'.
All following instructions are assumed to be executed from the aforementioned “live
filesystem” provided by the FreeBSD installation discs.
Before proceeding any further, the user is strongly urged to back up all data on the media
and the devices in question.
Furthermore, it will be assumed that the hard disk to be encrypted is accessible
through the device node /dev/ad0 and the removable (USB) medium through /dev/da0.
These paths must be adjusted to the actual setup!
3.4 Complete hard disk encryption using GBDE
3.4.1 Erasing previously stored data

Before a medium is set up to store encrypted data, it is important to completely erase all
data previously stored on it. All data on it has to be physically overwritten – ideally
multiple times. Otherwise the data that has previously been stored unencrypted would
still be accessible at the sector level of the hard disk until overwritten by new data. There
are two ways to wipe a hard disk clean:
# dd if=/dev/zero of=/dev/ad0 bs=1m
overwrites the entire hard disk space with zero values. The parameter bs sets the
block size to 1 MB – the default (512 B) would take a very long time with large disks.
# dd if=/dev/random of=/dev/ad0 bs=1m
does the same thing, but uses entropy instead of zero values to overwrite data. The
problem with the first approach is that it is quite obvious which parts of the medium
carry data and which ones are unused. Attackers looking for potential clues about the
encryption key can often exploit this information.
In most cases, however, this should not be a major risk. The downside of using
entropy is that it requires far more processing power than simply filling the hard disk
8
space with zero values. The required amount of time may therefore be too great a trade
off for the additional increase in security – especially on older, slower hardware.
3.4.2 Initialization & the lockfile

After the hard disk to be encrypted has been wiped clean, it can be initialized for
encryption. This is done using the gbde(8) command:
# gbde init /dev/ad0 -L /very/safe/place/lockfile
Enter new passphrase:
Reenter new passphrase:
The lockfile is very important, as it is needed later to access the master key which is
used to encrypt all data. The 16 bytes of data stored in this lockfile could also be saved in
the first sector of the medium or the partition, respectively. In that case, however, only
the passphrase would be required to get access to the data. The passphrase – however
strong it is – will face intensive exposure with mobile devices as it must be typed in each
time the system is booted up. It therefore cannot be realistically guaranteed that the
passphrase remains only known to those authorized to access the protected system and
data.
But since an additional medium is needed anyway in order to boot the core OS parts,
it might as well be used as a storage area for the lockfile – effectively functioning as a
kind of access token.
With this scheme, two things are required to get access to the data: the passphrase
and the lockfile. If the lockfile is unavailable (lost or destroyed), even knowledge of the
passphrase will not yield access to the data!
3.4.3 Attaching the encrypted medium

After the initialization is complete, the encrypted hard disk must now be attached –
meaning that the user has to provide both the passphrase and the lockfile to gbde,
which in turn provides (or denies) access to the decrypted data.
# gbde attach /dev/ad0 -l /very/safe/place/lockfile
Enter passphrase:
If the passphrase and the lockfile are valid, gbde creates an additional device node in
the /dev directory. This newly created node carries the name of the just attached device
('ad0') plus the suffix '.bde'.
• /dev/ad0 can be used to access the actual contents of the hard disk, in this case
the cipher text
• /dev/ad0.bde is an abstraction created by GBDE and allows plain text access to
the data
All reads from and writes to the .bdenode are automatically de/encrypted by GBDE
and therefore no user interaction is required once the correct passphrase and lockfile
has been provided.
The ad0.bde node acts just like the original ad0 node: it can be partitioned using
bsdlabel(8) or sliced with fdisk(8), it can be formated as well as mounted.
It is important to keep in mind that once a storage area has been attached and the
corresponding .bde device node for it has been created, it remains that way until it is
9
explicitly detached via the gbde command or the system is shut down. In the period
between attaching and detaching, there is no additional protection by GBDE.
3.4.4 Partitioning
The next step is to partition the hard disk. This is usually done using sysinstall(8) –
which, unfortunately, does not support GBDE partitions and fails to list device nodes
with a .bde suffix. Therefore, this work has to be done using the tool bsdlabel.
# bsdlabel -w /dev/ad0.bde
# bsdlabel -e /dev/ad0.bde
First, a standard label is written to the encrypted disk, so that it can be edited
afterwards. bsdlabel will display the current disk label in a text editor, so it can be
modified. In order to make the numbers in the following example easier to read, the disk
size is assumed to be 100 MB. The contents of the temporary file generated by bsdlabel
might look like this:
# /dev/ad0.bde:
8 partitions:
# size offset fstype [fsize bsize bps/cpg]
a: 198544 16 unused 0 0
c: 198560 0 unused 0 0 # "raw" part, don't edit
Each partition occupies one line. The values have the following meaning:
column description
1 a=boot partition; b=swap partition ; c=whole disk; d, e, f, g, h=freely available
2 and 3 partition size and its offset in sectors
4 filesystem type: 4.2BSD, swap or unused
5, 6 and 7 optional parameters, no changes required
Table 2: bsdlabel(8) file format
After the temporary file has been edited and the editor closed, bsdlabel will write
the label to the encrypted hard disk – provided no errors have been found (e.g.
overlapping partitions).
It is important to understand the device node names of the newly created partitions.
The encrypted boot partition (usually assigned the letter 'a'), is now accessible via device
node /dev/ad0.bdea. The swap partition is ad0.bdeb and so on. Just as adding a boot
partition to an unencrypted disk would result in a ad0a device node, adding an
encrypted slice holding several partitions inside would result in ad0s1.bdea, ad0s1.bdeb
and so on.
An easy way to keep the naming concept in mind is to remember that everything
written after the .bde suffix is encrypted and therefore hidden even to the kernel until the
device is attached.
For example: ad0s1.bdea means that the data on the first slice is encrypted –
including the information that there is a boot partition inside that slice. If the slice is not
attached, it is only possible to tell that there is a slice on the disk – neither the contents of
the slice, nor the fact that there is at least one partition inside the slice can be unveiled.
10
In fact, the node ad0s1.bdea does not even exist until the slice has been successfully
attached, because without having the key (and the lockfile), the kernel cannot know that
there is a partition inside the encrypted slide.
Scenario: multiple operating systems on the same disk
It is also possible to have multiple operating systems on the same disk – each on its
own slice. The slice containing FreeBSD can be encrypted completely, hiding even the
fact that the FreeBSD slice contains multiple partitions inside (boot, swap, etc). This way,
all data on the FreeBSD slice remains protected, while the other operating systems on
the machine can function normally on their unencrypted slices. In fact, they cannot even
compromise the data on the FreeBSD slice – even if an attacker manages to get root
access to a system residing on an unencrypted slice.
3.4.5 Creating the filesystem

Now that device nodes for the encrypted partitions exist, filesystems can be created on
them:
# newfs /dev/ad0.bdea
# newfs /dev/ad0.bded
etc.
Note that the swap partition does not need a filesystem; the 'c' partition represents
the entire (encrypted) disk. This partition must not be formated or otherwise be
modified!
3.4.6 Installing FreeBSD

Now that the filesystems have been created, FreeBSD can be installed on the encrypted
hard disk. Usually, this would be done using sysinstall again. But just as
sysinstall cannot partition and format encrypted media, it cannot install the system
on them. The distributions that comprise the FreeBSD operating system, therefore have
to be installed manually.
The FreeBSD installation disc contains a directory that is named after the release
version of the system, for example: 5.4RELEASE, 6.0BETA etc. In this directory, each
distribution – such as base, manpages or src – has its own subdirectory with an
install.sh script. The base distribution is required, all others are optional.
In order to install the distributions, the encrypted boot partition (and others, if used
for example for /usr) has to be mounted and the environment variable DESTDIR set to
the path where the encrypted boot partition has been mounted. Then all distributions
can be installed using their respective install.sh script.
The following example assumes that the encrypted boot partition /dev/ad0.bdea has
been mounted on /fixed and the FreeBSD installation disc on /dist (the “livefilesystem”
default). If the livefilesystem is used, the /fixed directory is easy to create because the
root (/) is a memory disk.
# mount /dev/ad0.bdea /fixed
# export DESTDIR=/fixed/
11
# cd /dist/5.4-RELEASE/base && ./install.sh
You are about to extract the base distribution into /fixed - are you SURE
you want to do this over your installed system (y/n)?
After all desired distributions have been installed, there is a complete FreeBSD
installation on the encrypted disk and the swap partition is also ready. But since this
system cannot be booted from the hard disk, it is necessary to set up the removable
medium.
3.4.7 Preparing the removable medium

As it has already been discussed, this medium will not be encrypted. This means that the
standard tool sysinstall can be used. The removable medium needs one partition of
at least 7 MB. This provides only space for the kernel, some modules and the utilities
required for mounting the encrypted partition(s). All other modules such as third party
drivers need to be loaded after init(8) has been invoked.
If it is desired that all FreeBSD kernel modules be available on the removable
medium and thus are loadable before init is called, the slice should be at least 25 MB in
size.
The removable medium can be sliced using fdisk or via 'Configure' 'Fdisk' in the
sysinstall menu. The changes made to the medium can be applied immediately by
hitting 'W'. After that, the slice has to be labeled (sysinstall menu 'Label'). All the
space on the slice can be used for the boot partition, since the swap partition on the
encrypted hard disk will be used. The mount point for the boot partition does not
matter; this text, however, will assume that it has been mounted on /removable.
sysinstall then creates the partition, the filesystem on it and also mounts it on the
specified location (/removable). After that, sysinstall can be quit in order to copy the
files required for booting from the removable medium. All that is required is the /boot
directory – it can be copied from the installation on the encrypted hard disk:
# cp -Rpv /fixed/boot /removable
3.4.8 The kernel modules

User interaction with GBDE is done through the userland tool gbde(8), but most of the
work is carried out by the kernel module geom_bde.ko. This module must be loaded
before the userland utility is called.
Usually, kernel modules are loaded by loader(8) based on the contents of the file
/boot/loader.conf – then control is passed over to the kernel. In order to have the GBDE
module loaded before init is executed, it must be loaded in advance by loader. The
following instruction adds the GBDE kernel module to the loader configuration file on
the removable medium:
# echo geom_bde_load=\“YES\“>> /removable/boot/loader.conf
In case additional kernel modules are needed at boot time, they must be copied to
/boot/kernel/ and appropriate entries must be added to /boot/loader.conf (this file
overrides the defaults in /boot/defaults/loader.conf).
In order to save space on the removable medium and also to speed up loading, all
kernel modules and even the kernel itself can be gzipped.
# cd /removable/boot/kernel
12
# gzip kernel geom_bde.ko acpi.ko
Binary code compresses to about half of the original size and thus brings a noticeable
decrease in loading time. The modules which will not be used or later will be loaded
from the hard disk can be deleted from the removable medium.
It is important, however, that the code on the removable medium (kernel, modules,
etc) is kept in sync with the system on the hard disk.
3.4.9 The problem with GBDE

As discussed earlier, GBDE has been designed with the encryption of partitions and even
entire media in mind. Unfortunately, however, the geom_bde.ko module does not allow
the kernel to mount an encrypted partition as the root filesystem.
This is because the passphrase must be provided through the utility in user space –
even though the module obviously operates in kernel space. So, by the time the kernel
must mount the root filesystem, the user has not even had the possibility of providing
the passphrase and attaching the encrypted device.
There are two solutions to this problem:
• The kernel must be modified to allow mounting of an encrypted root filesystem
by asking for the passphrase in kernel space. This way, the device node which
gives access to the decrypted data (the .bde device node) would be available
before init is started and could be specified in /etc/fstab as the root file
system. The new facility – GELI – has implemented this scheme and therefore
makes it a lot easier than the second solution.
• The second solution is not really a solution, but more a “hack”, as the
shortcomings of GBDE are not solved but avoided. The only conclusion is
therefore that the root filesystem cannot be encrypted and that the filesystem(s)
on the hard disk – although encrypted – must be mounted on directories
residing in the unencrypted root filesystem. Attaching and mounting the
encrypted hard disk must be done after the kernel has mounted an unencrypted
root filesystem and started init and subsequently gbde from it.
3.4.10 The memory disk

Since the contents of the root filesystem will not be encrypted, it is best to store on it
only what is needed to mount the encrypted partitions. Mounting the filesystem on the
removable medium as the root filesystem means that the removable medium would
have to be attached to the computer while the system is in use and therefore face a lot of
unnecessary exposure.
The better solution is to store an image of a memory disk on the removable medium,
which contains just the utilities necessary to mount the encrypted hard disk. The kernel
can mount the memory disk as the root filesystem and invoke init on it, so that gbde
can be executed. After the user has provided the passphrase to the encrypted partitions
on the hard disk, the utilities on the memory disk can mount the encrypted partitions
and then load the rest of the operating system from the encrypted hard disk – including
all applications and user data.
First, an image for the memory disk must be created on the removable medium.
# dd if=/dev/zero of=/removable/boot/mfsroot bs=1m count=10
13
Then a device node for the image is needed, so that a filesystem can be created on it
and then mounted.
# mdconfig -a -t vnode -f /removable/boot/mfsroot
md1
# newfs /dev/md1
# mount /dev/md1 /memdisk
If the output of mdconfig(8) differs from 'md1', the path in the following
instructions must be adjusted. The assumed mounting point for the memory disk will be
/memdisk.
3.4.11 Populating the memory disk filesystem

Since this filesystem is going to be mounted as the root filesystem, a directory must be
created to serve as the mount point for the encrypted boot partition (/memdisk/safe).
# cd /memdisk
# mkdir safe
Some other directories also act as mount points and do not need to be symlinked to
the encrypted hard disk. The directory /etc, however, is required, because the rc(8)
script in it will be modified to mount the encrypted partitions.
# mkdir cdrom dev dist mnt etc
Now, the lockfile, which is needed to access the encrypted data, must be copied onto
the removable medium – turning it into a kind of access token, without which the
encrypted data cannot be accessed even with the passphrase available.
# cp /very/safe/place/lockfile /memdisk/etc/
It is important to remember that the lockfile is updated each time the passphrase is
changed.
3.4.12 The booting process

After the kernel has been loaded from the removable medium it mounts the memory
disk as the root filesystem and then executes init, the first process. init in turn calls
rc, a script that controls the automatic boot process. Since rc is a text file rather than a
binary executable, it can be easily modified to mount the encrypted boot partition before
the majority of the system startup – which requires a lot of files – takes place. The rc
script can therefore be copied from the installation on the hard disk and then be edited.
# cp /fixed/etc/rc /memdisk/etc/
The following commands have to be inserted after the line “export HOME PATH” (in
5.4RELEASE: line 51) into /memdisk/etc/rc:
/rescue/gbde attach /dev/ad0 l /etc/lockfile && \
/rescue/mount /dev/ad0.bdea /safe && \
/rescue/mount w f /dev/md0 / && \
/rescue/rm R /etc && \
/rescue/ln s safe/etc /etc
14
The commands first attach the encrypted boot partition, mount it on /safe and then
erase the /etc directory from the memory disk, so that it can be symlinked to the
directory on the encrypted disk.
Obviously, the utilities in the /rescue directory need to be on the memory disk. The
/rescue directory is already part of a FreeBSD default installation and provides statically
linked executables of the most important tools. Although the size of the /rescue directory
seems at first glance to be huge (~470 MB!), there is in fact one binary which has been
hardlinked to the various names of the utilities. The /rescue directory therefore contains
about 130 tools which can be executed without any dependencies on libraries. The total
size is less than 4 MB. Although this fits easily on the created memory disk, the directory
cannot be just copied. The following example uses tar(1) in order to preserve the
hardlinks.
# cd /fixed
# tar -cvf tmp.tar rescue
# cd /memdisk
# tar -xvf /fixed/tmp.tar
# rm /fixed/tmp.tar
3.4.13 Creating the symlinks

The files required for mounting the encrypted boot partition are now in place and the rc
script has also been appropriately modified. But since the encrypted boot partition will
not be mounted as the root (/), but in a subdirectory of the memory disk (/safe), all of the
relevant directories must have entries in the root pointing to the actual directories in
/safe.
# umount /fixed
# mount /dev/ad0.bdea /memdisk/safe
# cd /memdisk
# ln -s safe/* .
3.4.14 Integrating the memory disk image

The memory disk image now contains all the necessary data, so it can be unmounted
and detached (if the memory disk image was not previously accessible through
/dev/md1, the third line must be adjusted).
# umount /memdisk/safe
# umount /memdisk
# mdconfig -d -u1
In order to save space and to speed up the booting process, the memory disk image
can also be gzipped, just like the kernel modules and the kernel itself:
# gzip /removable/boot/mfsroot
If the kernel was compiled with the MD_ROOT option – which is the case with the
GENERIC kernel – it is able to mount the root from a memory disk. The file that holds the
image of the memory disk must be loaded by the FreeBSD loader. This works almost
the same way as with kernel modules, as the image must be listed in the configuration
file /boot/loader.conf. Compared to executable code however, the memory disk image
15
must be explicitly specified as such in the configuration file, so the kernel knows how to
handle the file's contents. The following three lines are required in /boot/loader.conf on
the removable medium:
mfsroot_load="YES"
mfsroot_type="mfs_root"
mfsroot_name="/boot/mfsroot"
It is also important to note that there is no need to maintain an extra copy of the
/etc/fstab file on the removable medium as the kernel automatically mounts the first
memory disk that has been preloaded. Although this /etc/fstab issue is not a major
problem, it is a necessary measure in order to make this scheme work with GELI – which
is able to mount an encrypted partition as the root filesystem.
3.4.15 The swap partition

Although the swap partition has already been set up and is ready for use, the operating
system does not yet know which device to use. It is therefore necessary to create an entry
for it in the file /etc/fstab. This file must be stored on the hard disk, not the removable
medium.
# mount /dev/ad0.bdea /fixed/
# echo “/dev/ad0.bdeb none swap sw 0 0“ > /fixed/etc/fstab
Now, the system is finally ready and can be used by booting from the removable
medium. The modified rc script will ask for the passphrase and then mount the
encrypted partition, so that the rest of the system can be loaded.
3.4.16 Post-installation issues

Since the system on the encrypted disk was not installed using sysinstall, a few
things such as setting the timezone, the keyboard map and the root password have not
yet been taken care of. These settings can easily be changed by calling sysinstall now.
Packages such as the X server, which is not part of the system, can be added using
pkg_add(8). The system is now fully functional and ready for use.
3.5 Complete hard disk encryption using GELI

This chapter describes the process of setting up complete hard disk encryption using
FreeBSD's new GELI facility. GELI is so far only available on the 6.x branch. It is
important to note that the memory disk approach as discussed previously with GBDE is
also possible with GELI. But since GELI makes it possible to mount an encrypted
partition as the root filesystem, the memory disk is not a requirement anymore. This
advantage, however, is somewhat weakened by a drawback that the memory disk
scheme does not suffer from. This particular issue will be discussed in more detail later
and ultimately it is up to the user to decide which scheme is more appropriate.
As the concept of having a memory disk with GELI is very similar to having one with
GBDE, this chapter discusses only how to use GELI to boot directly with an encrypted
16
root filesystem – without the need for a memory disk.
Many of the steps required to make complete hard disk encryption work with GBDE
are also necessary with GELI – regardless of whether a memory disk is used or not.
Therefore the description and explanation of some steps will be shortened or omitted
completely here. The necessary commands will of course be given, but for a more
detailed explanation the respective chapters in the GBDE part are recommended for
reference.
3.5.1 Readying the hard disk

As it has already been mentioned in the GBDE chapter, erasure of previously stored data
on the medium intended for encryption is strongly recommended. The data can be
overwritten by either using the zero or the entropy device as a source.
# dd if=/dev/zero of=/dev/ad0 bs=1m
or
# dd if=/dev/random of=/dev/ad0 bs=1m
Their respective advantages and drawbacks were discussed in chapter 3.4.1.
3.5.2 Improvements and new problems with GELI

Just as GBDE, GELI must first initialize the medium intended for encryption. GELI's big
advantage over GBDE for the purpose of complete hard disk encryption is that it enables
the kernel to mount an encrypted partition as the root filesystem. This works by passing
the b parameter to the geli(8) userland tool when the hard disk is initialized. This
parameter causes GELI to flag the partition as “ask for passphrase upon discovery”.
When the kernel initializes the various storage media in the system at boot time, it
searches the partitions on them for any that have been flagged by the user and then asks
for the passphrase of the respective partition. The most important fact is, that this is
done in kernel space – the new device node providing access to the plain text (with the
suffix .eli, analogous to GBDE's .bde suffix) therefore already exists before the kernel
mounts the root filesystem.
Furthermore – as it is possible with GBDE – GELI also allows the key material to be
retrieved from additional sources besides the passphrase. While GBDE uses the 16byte
lockfile for this purpose, GELI supports the specification of a keyfile with the K
parameter. The size of this keyfile is not hardcoded into GELI and can be chosen freely
by the user; if '–' instead of a file name is given, GELI will read the contents of the keyfile
from the standard input.
This way it is even possible to concatenate several files and feed them to GELI's
standard input through a pipeline. The individual files would then each hold a part of
the key and the key would therefore be distributed across several (physical, if chosen)
places.
Unfortunately, however, the keyfile cannot be used with partitions which have been
flagged for “ask for passphrase upon discovery”. Using a passphrase and a keyfile to
grant access to the encrypted data would require that a parameter be passed to the
kernel – specifying the path to the keyfile. This path could of course also be hardcoded
into the kernel, for example that the keyfile must be located at /boot/geli.keys/<device>.
Unfortunately, this functionality does not yet exist in GELI. The ability to mount an
17
encrypted partition as the root filesystem comes therefore at the price of having to rely
only on the passphrase to protect the data. The memory disk approach that was
discussed in order to make complete hard disk encryption work with GBDE also works
with GELI. Although it is harder to set up and maintain, it combines the advantages of
“something you know” and “something you have”, namely a passphrase and a
lockfile/keyfile. Especially on mobile devices it is risky to rely only on a passphrase, since
it will face intensive exposure as it must be typed in each time the system is booted up.
The choice between better usability and increased security is therefore left to user.
3.5.3 Initializing, attaching and partitioning

Initializing the hard disk with GELI works similarly as it does with GBDE – except that the
partition must be flagged as “ask for passphrase upon discovery” and therefore cannot
(yet) use a keyfile.
# geli init -b /dev/ad0
Enter new passphrase:
Reenter new passphrase:
Very important here is to specify the b parameter, which causes the geom_eli.ko
kernel module to ask for the passphrase if a GELI encrypted partition has been found.
The a parameter can (optionally) be used to specify the encryption algorithm: AES,
Blowfish or 3DES.
If this setup is performed directly from the 'fixit' live filesystem, then the /lib
directory must be created by symlinking it to the existing /dist/lib directory. This is
necessary because GELI needs to find its libraries in /lib. The GELI executable will
actually run without /lib, but will then hide its features from the user – therefore making
the problem much less obvious.
Attaching the hard disk is also largely the same as with GBDE, again except that the
keyfile parameter must be omitted from the command.
# geli attach /dev/ad0
Enter passphrase:
Upon successful attachment, a new device node will be created in the /dev directory
which carries the name of the specified device plus a '.eli' suffix. Just like the '.bde'
device node created by GBDE, this node provides access to the plain text. The output of
geli after successful attachment looks something like this (details depend on the
parameters used and the available hardware):
GEOM_ELI: Device ad0.eli created.
GEOM_ELI: Cipher: AES
GEOM_ELI: Key length: 128
GEOM_ELI: Crypto: software
Since sysinstall cannot read GELI encrypted partitions either, the partitioning
must be done using the bsdlabel tool.
# bsdlabel -w /dev/ad0.eli
# bsdlabel -e /dev/ad0.eli
Partition management was discussed in more detail in chapter 3.4.4.
18
3.5.4 Filesystem creation and system installation
Now that the partition layout has been set, the filesystem(s) can be created, so FreeBSD
can be installed.
# newfs /dev/ad0.elia
# newfs /dev/ad0.elid
etc.
The actual installation of the system on the encrypted hard disk must also be done
manually, since sysinstall does not support GELI encrypted partitions.
# mount /dev/ad0.elia /fixed
# export DESTDIR=/fixed/
# cd /dist/6.0-RELEASE/base && ./install.sh
You are about to extract the base distribution into /fixed - are you SURE
you want to do this over your installed system (y/n)?
3.5.5 The removable medium

Since this medium is not going to be encrypted, it can be sliced and partitioned with
sysinstall. The size requirements are largely the same as for GBDE – the minimum is
even a bit lower because there is no need to store the image of the memory disk. With a
customized kernel, this minimum may be as low as 4 MB.
In order to boot the kernel from the removable medium (/removable), it is necessary
to copy the /boot directory from the encrypted hard disk (mounted on /fixed).
# cp -Rpv /fixed/boot /removable
All kernel modules except geom_eli.ko and its dependency crypto.ko (and acpi.ko, if
used) can be deleted if space is a problem. Further, all modules and even the kernel can
be gzipped. This saves not only space, but also reduces loading time.
# cd /removable/boot/kernel
# gzip kernel geom_eli.ko acpi.ko
Just as it is the case with GBDE, GELI also needs its kernel module geom_eli.ko
loaded by loader(8) in order to ask for the passphrase before the root filesystem is
mounted. The following command adds the appropriate entry to /boot/loader.conf.
# echo geom_eli_load=\“YES\“>> /removable/boot/loader.conf
3.5.6 Mounting the encrypted partition

Because of GELI's ability to mount encrypted partitions as the root filesystem the entire
workaround with the memory disk can be avoided. So far, however, the kernel does not
know which partition it must mount as the root filesystem – even if the device node to
the plain text of the encrypted hard disk has been created by GELI. The memory disk
approach, which is necessary to make complete hard disk encryption work with GBDE,
has the advantage that the kernel will automatically mount the memory disk as the root
filesystem if an image has been preloaded.
In this case, however, it is necessary to create an entry in /etc/fstab, so the kernel
knows which partition to mount as the root filesystem.
# mkdir /removable/etc
19
# echo “/dev/ad0.elia / ufs rw 1 1” >> /removable/etc/fstab
It is important to note that this file must be stored on the removable medium and
serves only the purpose of specifying the device for the root filesystem. As soon as the
kernel has read out the contents of the file, it will mount the specified device as the root
filesystem and the files on the removable medium (including fstab) will be outside of the
filesystem name space. This means that the removable medium must first be mounted
before the files on it can be accessed through the filesystem name space. It also means,
however, that the removable medium can actually be removed after the root filesystem
has been mounted from the encrypted hard disk – thus reducing unnecessary exposure.
It is crucial that the removable medium be always in the possession of the user, because
the whole concept of complete hard disk encryption relies on the assumption that the
boot medium – therefore the removable medium, not the hard disk – is uncompromised
and its contents are trusted.
If any other partitions need to be mounted in order to boot up the system – for
example /dev/ad0.elid for /usr – they must be specified in /etc/fstab as well. Since most
installations use at least one swap partition, the command for adding the appropriate
entry to /etc/fstab is given below.
# echo “/dev/ad0.elib none swap sw 0 0“ > /fixed/etc/fstab
The system is now ready for use and can be booted from the removable medium. As
the different storage devices in the system are found, GELI searches them for any
partitions that were initialized with the geli init -b parameter and asks for the
passphrase. If the correct one has been provided, GELI will create new device nodes for
plain text access to the hard disk and the partitions on it (e.g. /dev/ad0.elia), which then
can then be mounted as specified in /etc/fstab.
After that, the rest of the system is loaded. sysinstall can then be used in order to
adjust the various settings that could not be set during the installation procedure – such
as timezone, keyboard map and especially the root password!
4 Complete hard disk encryption in context
4.1 New defenses & new attack vectors – again

Any user seriously thinking about using complete hard disk encryption should be aware
of what it actually protects and what it does not.
Since encryption requires a lot of processing power and can therefore have a
noticeable impact on performance, it is usually not enabled by default. FreeBSD marks
no exception here. Although it provides strong encryption algorithms and two powerful
tools for encrypting storage media, it is up to the user to discover and apply this
functionality.
This paper gave instructions on how to encrypt an entire hard disk while most of the
operating system is still stored and loaded from it. It is important to remember, however,
that FreeBSD – or any other software component for that matter – will not warn the user
if the encrypted data on the hard disk is leaked (see chapter 2.3) or intentionally copied
to another, unencrypted medium, such as an external drive or a smart media card. It is
the responsibility of the user to encrypt these media as well.
This responsibility applies equally well to data in transit. Network transmissions are
20
in most cases not encrypted by default either. Since all encryption and decryption of the
data on the hard disk is done transparently to the user once the passphrase has been
provided, it is easy to forget that some directories might contain data which is stored on
a different machine and made available through NFS, for example – in which case the
data is transferred in the clear over the network, unless explicitly set up otherwise.
The mounting facility in UNIX is very powerful; but it also makes it difficult to keep
track of which medium actually holds what data.
The network poses of course an additional threat, because of an attacker's ability to
target the machine remotely. The problem has already been discussed in chapter 1. If a
particular machine is easier to attack remotely than locally, any reasonable attacker will
not even bother with getting physical access to the machine. In that case it would make
no sense to use complete hard disk encryption, because it does not eliminate the
weakest link (the network connectivity).
If, on the other hand, not the network, but the unencrypted or not fully encrypted
hard disk is the weakest link and the attacker is also capable of getting physical access to
the machine (for reasons discussed in chapter 2.4), then complete hard disk encryption
makes sense.
A key point to remember is that as long as a particular storage area is attached, the
data residing on it is not protected any more than any other data accessible to the
system. This applies to both GBDE and GELI; even unmounting an encrypted storage
area will not protect the data from compromise since the corresponding device node
providing access to the plain text still exists. In order to remove this plain text device
node, the storage area in question must be detached. With GBDE this must be done
manually, GELI has a feature that allows for automatic detachment on the last close –
but this option must be explicitly specified.
Since the partition holding the operating system must always be attached and
mounted, its contents are also vulnerable during the entire time the system is up. This
means that remotely or even locally introduced viruses, worms and trojans can
compromise the system in the same way they can do it on a system without complete
hard disk encryption.
Another way to attack the system would be by compromising the hardware itself, for
example by installing a hardware keylogger. This kind of attack is very hard to defend
against and this paper makes no attempt to solve this issue.
What complete hard disk encryption does protect against, is attacks which aim at
either accessing data by reading out the contents of the hard disk on a different system
in order to defeat the defenses on the original system or by compromising the system
stored on the hard disk, so the encryption key or the data itself can be leaked. Encryption
does not, however, prevent the data from being destroyed, both accidentally and
intentionally.
If it is chosen that the encrypted partition is mounted directly as the root filesystem –
without the need for a memory disk, then it is crucial that a strong passphrase be
chosen, because that will be the only thing required to access the encrypted data.
Choosing the memory disk approach makes for a more resilient security model, since it
enables the user to use a lockfile (GBDE) or a keyfile (GELI) – in order to get access to the
data.
While all these previously mentioned conditions and precautions matter, it is
absolutely crucial to understand that the concept of complete hard disk encryption
depends upon the assumption that the data on the removable medium is trusted.
The removable medium must be used because the majority of the hardware is not
capable of booting encrypted code. Since the kernel and all the other code necessary for
mounting the encrypted partition(s) must be stored in the clear on the removable
21
medium, the problem of critical code getting compromised has, in fact, not really been
solved. The most efficient way to attack a system like this would most likely be by
compromising the code on the removable medium.
It is therefore crucial that the user keep the removable medium with him or her at all
times. If there is the slightest reason to believe that the data on it may have been
compromised, its contents must be erased and reconstructed according to the instructions
in the respective GBDE or GELI chapters.
If the removable medium has been lost or stolen and there was a keyfile or lockfile
stored on it, then two issue must be taken into account:
• The user will not be able to access to encrypted data even with the passphrase.
It is therefore strongly recommended that a backup of the keyfile/lockfile be
made and kept in a secure place – preferably without network connectivity.
• The second possibility can be equally devastating, since the keyfile/lockfile
could fall into the hands of someone who is determined to break into the
system. In that case, all the attacker needs is the passphrase – which can be very
hard to keep secret for a mobile device. It is therefore recommended that both
the passphrase and the keyfile/lockfile are changed in the event of a removable
medium loss or theft.
4.2 Trade-offs
Complete hard disk encryption offers protection against specific attacks as discussed in
chapter 4.1. This additional protection, however, comes at a cost – which is usually why
security measures are not enabled by default. In the case of complete hard disk
encryption, the tradeoffs worth mentioning the most are the following:
• Performance. Encryption and decryption consume a lot of processing power.
Since each I/O operation on the encrypted hard disk requires additional
computation, the throughput is often limited by the power of the CPU(s) and
not the bandwidth of the storage medium. Especially write operations, which
must be encrypted, are noticeably slower than read operations, where
decryption is performed. Systems which must frequently swap out data to
secondary storage and therefore usually to the encrypted hard disk can suffer
from an enormous performance penalty. In cases where performance becomes
too big a problem it is suggested that dedicated hardware be used for
cryptographic operations. GELI supports this by using the crypto(9) framework,
GBDE unfortunately does so far not allow for dedicated hardware to be used
and must therefore rely on the CPU(s) instead.
• Convenience. Each time the system is booted, the user is required to attach or
insert the removable medium and enter the passphrase. Booting off a
removable medium is usually slower than booting from a hard disk and the
passphrase introduces an additional delay.
• Administrative work. Obviously the whole scheme must first be set up before it
can be used. The majority of this quite lengthy process must also be repeated
with each system upgrade as the code on the removable medium must not get
out of sync with the code on the hard disk. As this setup or upgrade process is
also prone to errors such as typos, it may be considered an additional risk to the
data stored on the device.
22
This list is by no means exhaustive and every user thinking about using complete hard
disk encryption is strongly encouraged to carefully evaluate its benefits and drawbacks.
4.3 GBDE vs. GELI

FreeBSD provides two tools for encrypting partitions, GBDE and GELI. Both can be used
to make complete hard disk encryption work. If GBDE is chosen, the memory disk
approach must be used, as GBDE does not allow the kernel to mount an encrypted
partition as the root filesystem. The advantage is that it is possible to use a lockfile in
addition to a passphrase. This makes for a more robust security model and should
compensate for the administrative “overhead” caused by the memory disk.
GELI not only makes it possible to use a memory disk too, it also allows the user to
choose from different cryptographic algorithms and key lengths. In addition to that it
also offers support for dedicated cryptographic hardware devices and of course
eliminates the need for a memory disk by being able to directly mount the encrypted
boot partition. The drawback of mounting the root directly from an encrypted partition
is that GELI so far does not allow for a keyfile to be used and therefore the security of the
encrypted data depends solely on the passphrase chosen.
Looking at the features of the two tools, it may seem as though GELI would be the
better choice in any situation. It should be noted, however, that GBDE has been around
for much longer than GELI and therefore is more likely to have received more testing and
review.
5 Conclusion
Mobile devices are intended to be used anywhere and anytime. As these devices get
increasingly sophisticated, they allow the users to store massive amounts of data – a lot
of which may often be sensitive. Encrypting individual files simply does not scale and on
top of that does nothing to prevent the data from leaking to other places. Partitionbased
encryption scales much better but still, a lot of information can be compiled from
unencrypted sources such as system log files, temporary working copies of opened files
or the swap partition. In addition to that, both schemes do nothing to protect the
operating system or the applications from being compromised.
In order to defend against this kind of attack, it is necessary to encrypt the operating
system and the applications as well and boot the core parts such as the kernel from a
removable medium. Since the boot code must be stored unencrypted in order to be
loaded, it must be kept on a medium that can easily be looked after.
FreeBSD provides two tools capable of encrypting disks: GBDE and GELI. Complete
hard disk encryption can be accomplished by using either a memory disk as the root
filesystem and then mount the encrypted hard disk in a subdirectory or by directly
mounting the encrypted hard disk as the root filesystem.
The first approach can be done with both GBDE and GELI and has the advantage
that a lockfile or keyfile can be used in addition to the passphrase, therefore providing
more robust security. The second approach omits the memory disk and therefore saves
some administrative work. It works only with GELI, however, and does not allow for a
keyfile to be used – therefore requiring a tradeoff between better usability/maintain
ability and security.
23
Under no circumstances does complete hard disk encryption solve all problems
related to security or protect against any kind of attack. What it does protect against, is
attacks which are aimed at accessing data by reading out the contents of the particular
hard disk on a different system in order to defeat the original defenses or to compromise
the operating system or applications in order to leak the encryption key or the encrypted
data itself.
As with any security measure, complete hard disk encryption requires the users to
make tradeoffs. The increase in security comes at the cost of decreased performance,
less convenience and more administrative work.
Complete hard disk encryption makes sense if an unencrypted or partially encrypted
hard disk is the weakest link to a particular kind of attack.
References & further reading

Dawidek, 2005a
P. J. Dawidek, geli – control utility for cryptographic GEOM class
FreeBSD manual page
April 11, 2005
Dawidek, 2005b
P. J. Dawidek, GELI disk encryption GEOM class committed
http://lists.freebsd.org/pipermail/freebsdcurrent/2005July/053449.html
posted on the 'freebsdcurrent' mailing list
July 28, 2005
Dowdeswell & Ioannidis, 2003
R. C. Dowdeswell & J. Ioannidis, The CryptoGraphic Disk Driver
http://www.usenix.org/events/usenix03/tech/freenix03/full_papers/dowdeswell/
dowdeswell.pdf
June 2003
Kamp, 2003a
P.H. Kamp, GBDE – GEOM Based Disk Encryption
http://phk.freebsd.dk/pubs/bsdcon03.gbde.paper.pdf
July 7, 2003
Kamp, 2003b
P.H. Kamp, GEOM Tutorial
http://phk.freebsd.dk/pubs/bsdcon03.slides.geomtutorial.pdf
August 19, 2003
Lemos, 2005
R. Lemos, Backups tapes a backdoor for identity thieves
http://www.securityfocus.com/news/11048
April 28, 2005
Leyden, 2004
J. Leyden, Oops! Firm accidentally eBays customer database
http://www.theregister.co.uk/2004/06/07/hdd_wipe_shortcomings/
June 7, 2004
24
Noguchi, 2005
Y. Noguchi, Lost a BlackBerry? Data Could Open A Security Breach
http://www.washingtonpost.com/wpdyn/content/article/2005/07/24/
AR2005072401135.html
July 25, 2005
OpenBSD, 1993
vnconfig configure vnode disks for file swapping or pseudo file systems
OpenBSD manual page
http://www.openbsd.org/cgibin/man.cgi?query=vnconfig&sektion=8&arch=i386&
apropos=0&manpath=OpenBSD+Current
July 8, 1993
Reuters, 2005
Reuters, Stolen PCs contained Motorola staff records
http://news.zdnet.co.uk/internet/security/0,39020375,39203514,00.htm
June 13, 2005
Sarche, 2005
J. Sarche, Hackers hit U.S. Army computers,

http://www.globetechnology.com/ servlet/story/RTGAM.20050913.gtarmysep13/

BNStory/Technology/
September 13, 2005
25

Complete Hard Disk Encryption

Uploaded by

Complete Hard Disk Encryption

Uploaded by

Complete Hard Disk Encryption Using FreeBSD's

2 Partial disk encryption

2.1 File-based encryption

2.2 Partition-based encryption

2.3 The leakage risk

2.4 New attack vectors

3 Complete disk encryption

3.1 Tools provided by FreeBSD

3.2 The problem with complete disk encryption

3.4 Complete hard disk encryption using GBDE

3.4.1 Erasing previously stored data

3.4.2 Initialization & the lockfile

3.4.3 Attaching the encrypted medium

3.4.5 Creating the filesystem

3.4.6 Installing FreeBSD

3.4.7 Preparing the removable medium

3.4.8 The kernel modules

3.4.9 The problem with GBDE

3.4.10 The memory disk

3.4.11 Populating the memory disk filesystem

3.4.12 The booting process

3.4.13 Creating the symlinks

3.4.14 Integrating the memory disk image

3.4.15 The swap partition

3.4.16 Post-installation issues

3.5 Complete hard disk encryption using GELI

3.5.1 Readying the hard disk

3.5.2 Improvements and new problems with GELI

3.5.3 Initializing, attaching and partitioning

3.5.5 The removable medium

3.5.6 Mounting the encrypted partition

4 Complete hard disk encryption in context

4.1 New defenses & new attack vectors – again

4.3 GBDE vs. GELI

References & further reading

You might also like