dm-crypt: a device-mapper crypto target

About

Device-mapper is a new infrastructure in the Linux 2.6 kernel that provides a generic way to create virtual layers of block devices that can do different things on top of real block devices like striping, concatenation, mirroring, snapshotting, etc... The device-mapper is used by the LVM2 and EVMS 2.x tools.
dm-crypt is such a device-mapper target that provides transparent encryption of block devices using the new Linux 2.6 cryptoapi. The user can basically specify one of the symmetric ciphers, a key (of any allowed size), an iv generation mode and then he can create a new block device in /dev. Writes to this device will be encrypted and reads decrypted. You can mount your filesystem on it as usual. But without the key you can't access your data.
It does basically the same as cryptoloop only that it's a much cleaner code and better suits the need of a block device and has a more flexible configuration interface. The on-disk format is also compatible. In the future you will be able to specify other iv generation modes for enhanced security (you'll have to reencrypt your filesystem though).

I've set up a Wiki. It's naked at the moment, feel free to fill it with some useful informations.
There's a mailing list at dm-crypt@saout.de. If you want to subscribe just send an empty mail to dm-crypt-subscribe@saout.de.
Gmane provides a NNTP interface and web archive for this mailing list.

Download

There is support for dm-crypt in the latest official kernel 2.6.4 which you can find on kernel.org. Please use the mirrors for downloads.
There is a HIGHMEM cryptoapi bug in kernels before 2.6.4-rc2, please upgrade if you were using such a kernel.
The latest version of the native userspace setup tool is cryptsetup 0.1.
Clemens Fruhwirth is maintaining an enhanced version of cryptsetup with the LUKS extension that allows you to have an on-disk block of metadata which is superior to the current mechanism and was my long term plan anyway but I didn't find the time to implement that yet...

Description

NEW: I've set up a Wiki. It's naked at the moment, feel free to fill it with some useful informations.

Installation:
Once you have a Linux 2.6 kernel with dm-crypt support on your machine, you need to activate device-mapper and dm-crypt in your kernel.
You can find both config options under Device Drivers > Multi-device support (RAID and LVM). Both can be compiled statically or as modules (code which you can insert and remove from the kernel at runtime).
The config options are also called CONFIG_BLK_DEV_DM and CONFIG_DM_CRYPT.
You also need some userspace tools. You need to install the device-mapper package, you can find the latest version here.

If you have compiled device-mapper as a module you must load it using modprobe dm-mod, the dm-crypt module should autoload when used.
You should make sure that you have the /dev/mapper directory and the /dev/mapper/control device node.
If not, you should follow the instructions in the INSTALL file found in the device-mapper package. The INTRO file also explains some device-mapper basics which might be useful.

Setup:
The mapped device can be created through userspace tools calling the appropriate device-mapper ioctl. Since there are no dedicated tools yet everything is done through dmsetup.
(note: If you don't want to know the details you might want to skipt the next paragraphs and directly go to the description of cryptsetup)

dmsetup is used to create and remove devices, get information about devices or reload tables (that means changing the mapping while the device is in use).
The syntax for device creation is: dmsetup create <name>
<name> is the name of the created device. It will appear under /dev/mapper/<name>.
dmsetup then expects the table on stdin (you could also give a file name as third parameter).
The table is a list of lines with a sector range, target type and target config. It looks like:
<start sector> <sector count> <target type> <arguments>
I'm not going into every detail here. A dm-crypt table looks like:

0 <sector count> crypt <sector format> <key> <IV offset> <real device> <sector offset>

<sector format>
It consists of the name of the symmetric encryption cipher and an optional IV generation mode (if none is selected the cryptoloop compatible plain mode is chosen).
Currently the IV generation modes ecb and plain are available. ecb means no cipher block chaining and IV is used at all (don't do this if you don't know what you're doing, it is a security catastrophe!). plain uses the unhashed 32 bit sector number as IV. The IV generation mode is appended using a dash or since Linux 2.6.10 cbc-essiv:<hash> (this prevents a watermark attack weakness). Examples for this parameter are: des, aes-cbc-essiv:sha256 or twofish-ecb.
cat /proc/crypto will show you the supported ciphers.
<key>
This is the key used for encryption. It is encoded as a hexadecimal number. You can only use key sizes that are valid for the selected cipher. See /proc/crypto output.
<IV offset>
The IV offset is a sector count that is added to the sector number before creating the IV. It can be used to create a map that starts after the first encrypted sector. Usually you'll set it to zero except your device is only partially available or something like this.
<real device>
This is the device that is going to be used as backend and contains the encrypted data. You can specify the path like /dev/xxx or a device number <major>:<minor>.
<sector offset>
Finally this is the sector (number of 512 byte sectors) offset where the encrypted data begins in the given real device.

So a complete line to setup the device might look like:
echo 0 `blockdev --getsize /dev/hda5` crypt aes-plain 0123456789abcdef0123456789abcdef 0 /dev/hda5 0 | dmsetup create volume1
Note the use of the blockdev command to get the number of sectors on /dev/hda5.
The created device will be named /dev/mapper/volume1.

The device can then be mounted (you should not forget to create a filesystem first). You can remove the device again using dmsetup remove <name>. If the creation fails see the syslog for kernel messages. Don't forget to remove the device before trying to recreate it.

Except for the additional parameters dmsetup can be used somewhat like losetup for cryptoloop. You can use hexdump to create the hex key representation and pipe the output from hashalot into it or something. This is what the cryptsetup tool below does.

cryptsetup:
Because the way using dmsetup directly is too complicated for most people I'm currently writing a native cryptsetup program to behave like one of the patched losetup's out there. It's going to support a lot more features in the future.
NEW: A first version of the native cryptsetup implementation in C is ready:
cryptsetup 0.1 requires libgcrypt 1.1.x and libdevmapper

I've got a CVS server, if you're interested in development I can give you write access. The old version is a shell script that uses dmsetup and hashalot to do the same. I'm using the same syntax. The script can be found here. You can put it into /usr/local/sbin or somewhere you like. The old script requires the tools dmsetup, hashalot, hexdump, sed, head, awk and ls. Most of these will most likely come with your distribution.
The hashalot tool can be found here (not required with -h plain).
The dmsetup tool can be found in the device-mapper package here.
Don't forget to call scripts/devmap_mknod.sh (only once) in the device-mapper package to create the /dev/mapper/control device node if you don't use devfs or udev.

Syntax: /usr/local/sbin/cryptsetup [<OPTIONS>] <action> <name> [<device>]
        <OPTIONS>:
            -c <cipher> (see /proc/crypto)
            -h {plain/<hash>} (see hashalot, WARNING: use ripemd160 instead of rmd160)
            -y (verifies the passphrase by asking for it twice)
            -d <file> (read key from file
                 e.g. /dev/urandom; useful for swap devices.
                 If set, the parameters -h and -y will be ignored)
            -s <keysize> (in bits) (WARNING: in bytes for cryptsetup.sh)
            -b <size> (in sectors)
            -o <offset> (in sectors)
            -p <skipped> (in sectors)
        <action> is one of:
            create - create device
            remove - remove device
            reload - modify active device
            resize - resize active device
            status - show device status
        <name> is the device to create under /dev/mapper/
        <device> is the encrypted device

When creating a device the program will ask for the a passphrase. The passphrase will then be hashed using the hashalot program and be used as key.
Alternatively a passphrase can be piped through stdin.
The hashing can be turned off with -h plain.

The defaults are aes with a 256 bit key, hashed using ripemd160. Since Linux 2.6.10 you can use an alternative IV scheme to prevent a watermark attack weakness. aes-cbc-essiv:sha256 should do it.

Don't forget: cryptsetup only creates a mapping. If you call cryptsetup again after a reboot and supply the same passphrase you will be able to mount your filesystem you created before.

Migration from cryptoloop and compatibility

The on-disk layouts used by the current 2.6 cryptoloop are supported by dm-crypt.
Cryptoloop also uses cryptoapi so the name of the ciphers are the same. Cryptoloop also supports ECB and CBC mode. Use <cipher>-ecb and <cipher>-plain accordingly with dm-crypt. If you didn't explicitly specify either -ecb or -cbc before you don't need it now, the default plain IV generation will be used. There will be additional (incompatible, but more secure) possibilites in the future because the unhashed sector number as IV is too predictible.

You'll need to figure out how your passphrase was turned into a key to use for losetup. There are several patches floating around doing things differently. But usually cryptsetup will provide a working solution to recreate the same key from your passphrase.

If you want to migrate from 2.4 cryptoloop please take a look at Clemens Fruhwirth's Cryptoloop Migration Guide. He describes the differences between 2.4 and 2.6 cryptoapi (or basically the bugs in 2.4 cryptoapi...). If you need to cut the key size you can use the -s option instead of playing with dd.
(BTW: Clemens has a i586 optimized version of the aes and serpent cipher on his page, about twice as fast as the kernel implementation.)

Why

Why dm-crypt?
Originally it started as a fun project because I wanted to play with the new Linux 2.6 internals. I got a lot of great help from the device-mapper guys at Sistina (now Redhat). Thank you very much!
It turned out that this implementation worked great and is very clean compared to the hacked loop device. The device-mapper core provides much better facilities to stack block devices. dm-crypt uses mempools to assure we never run into out-of-memory deadlocks when allocating buffers.
Also the device-mapper configuration interface provides much more flexibility than the losetup ioctl. And you can create as many devices as you want with any names you want and combine them with other dm targets. Online device resizing is also possible, e.g. if you use dm-crypt on top of a logical volume. There might perhaps even be LVM or EVMS support for device encryption in the future.

FAQ

But I don't want to use LVM!
You don't need LVM. Device-mapper is an all-purpose kernel feature, not tied to LVM in any way.

What if I want to encrypt a filesystem and keep it in a file?
You can use dm-crypt on top of a normal loop device, call losetup and cryptsetup.
I'm going to add loop support to cryptsetup so it can do this for you.

I created my filesystem on the encrypted device. How can I keep it across reboots?
Very simple. Call cryptsetup again and supply the same passphrase. It only creates a mapping, not a filesystem.

What if I want to change my passphrase?
At the moment you'll need to reencrypt your device because the passphrase is directly tied to the key.
There are plans to write a tool that stores the master key on disk and encrypted so it can be unlocked using a passphrase. You can then change your passphrase on a regular basis.
If you want to reencrypt your filesystem you'll have to recreate a new one and move your files.
(I've got an experimantal tool in the works that allows you to reencrypt your block device on the fly, assuming you don't reboot your machine...)

I've read about security problems.
Yes, the IV schemes currently supported by dm-crypt are the same as the ones supported by cryptloop. There's the ECB mode which is a catastrophe (no IV at all) and the "plain" mode, which is already a lot better. Older cryptoloops used ECB by default, but with dm-crypt the default is "plain" (which is the unhashes sector number used as IV).
Since dm-crypt is extensible there will be better possibilities in the future, but they will be on-disk incompatible with cryptoloop so you'll have to reencrypt.

Help! I can't figure out how to use my old encrypted data! I was using...
There are different implementations out there. Some are non-cryptoapi and/or broken implementations. SuSE uses its own loop-twofish implementation which makes dangerous assumptions and is broken when changing the blocksize ("timebomb crypto"). You cannot use this with dm-crypt.

Can I reencrypt my data without copying all the files?
There's an experimental and unfinished dmconvert program that can reencrypt the data while the filesystem is mounted. If you can get it running it should be safe enough to not eat your data, but make sure you don't interrupt it or crash your system while it is running. Don't blame me if something goes wrong.

Can I use encrypted swap?
Yes. You can specify a key file /dev/random and run mkswap afterwards, so the device will be created with a different key each time and the data is not accessible at all after a reboot.

Is there a mailing list?
There's a mailing list at dm-crypt@saout.de. If you want to subscribe just send an empty mail to dm-crypt-subscribe@saout.de.
Gmane provides a NNTP interface and web archive for this mailing list.

My system hangs for some time in regular intervals when writing to encrypted disks.
You are probably using Linux 2.6.4. Du to the introduction of kthread pdflush is running at nice level -10, which means that the kernels treats dm-crypt writes as a real time task and doesn't allow scheduling.
Solution: Switch to 2.6.5 or later or renice pdflush manually.

Can I use the mount command itself to do all the magic needed?
I've written an experimental patch for this, see my post in the mailing list archive.

Where can I send my contributions?
Because maintaining a web page takes time and people keep mailing me a lot of things I could integrate they can enter it into this nice Wiki.

Questions, suggestions, criticism?

Feel free to contact me: christophe@saout.de.

Valid XHTML 1.0! Valid CSS!