Frequently Asked Questions Cryptsetup/LUKS
Sections
1. General Questions
2. Setup
3. Common Problems
4. Troubleshooting
5. Security Aspects
6. Backup and Data Recovery
7. Interoperability with other Disk Encryption Tools
8. Issues with Specific Versions of cryptsetup
9. The Initrd question
10. LUKS2 Questions
11. References and Further Reading
A. Contributors
1. General Questions
* 1.1 What is this?
This is the FAQ (Frequently Asked Questions) for cryptsetup. It covers
Linux disk encryption with plain dm-crypt (one passphrase, no
management, no metadata on disk) and LUKS (multiple user keys with one
master key, anti-forensic features, metadata block at start of device,
...). The latest version of this FAQ should usually be available at
https://gitlab.com/cryptsetup/cryptsetup/wikis/FrequentlyAskedQuestions
* 1.2 WARNINGS
LUKS2 COMPATIBILITY: This FAQ was originally written for LUKS1, not
LUKS2. Hence regarding LUKS2, some of the answers found here may not
apply. Updates for LUKS2 have been done and anything not applying to
LUKS2 should clearly say LUKS1. However, this is a Frequently Asked
Questions, and questions for LUKS2 are limited at this time or at least
those that have reached me are. In the following, "LUKS" refers to both
LUKS1 and LUKS2.
The LUKS1 on-disk format specification is at
https://www.kernel.org/pub/linux/utils/cryptsetup/LUKS_docs/on-disk-format.pdf
The LUKS2 on-disk format specification is at
https://gitlab.com/cryptsetup/LUKS2-docs
ATTENTION: If you are going to read just one thing, make it the section
on Backup and Data Recovery. By far the most questions on the
cryptsetup mailing list are from people that managed to damage the start
of their LUKS partitions, i.e. the LUKS header. In most cases, there
is nothing that can be done to help these poor souls recover their data.
Make sure you understand the problem and limitations imposed by the LUKS
security model BEFORE you face such a disaster! In particular, make
sure you have a current header backup before doing any potentially
dangerous operations. The LUKS2 header should be a bit more resilient
as critical data starts later and is stored twice, but you can decidely
still destroy it or a keyslot permanently by accident.
DEBUG COMMANDS: While the --debug and --debug-json options should not
leak secret data, "strace" and the like can leak your full passphrase.
Do not post an strace output with the correct passphrase to a
mailing-list or online! See Item 4.5 for more explanation.
SSDs/FLASH DRIVES: SSDs and Flash are different. Currently it is
unclear how to get LUKS or plain dm-crypt to run on them with the full
set of security assurances intact. This may or may not be a problem,
depending on the attacker model. See Section 5.19.
BACKUP: Yes, encrypted disks die, just as normal ones do. A full backup
is mandatory, see Section "6. Backup and Data Recovery" on options for
doing encrypted backup.
CLONING/IMAGING: If you clone or image a LUKS container, you make a copy
of the LUKS header and the master key will stay the same! That means
that if you distribute an image to several machines, the same master key
will be used on all of them, regardless of whether you change the
passphrases. Do NOT do this! If you do, a root-user on any of the
machines with a mapped (decrypted) container or a passphrase on that
machine can decrypt all other copies, breaking security. See also Item
6.15.
DISTRIBUTION INSTALLERS: Some distribution installers offer to create
LUKS containers in a way that can be mistaken as activation of an
existing container. Creating a new LUKS container on top of an existing
one leads to permanent, complete and irreversible data loss. It is
strongly recommended to only use distribution installers after a
complete backup of all LUKS containers has been made.
UBUNTU INSTALLER: In particular the Ubuntu installer seems to be quite
willing to kill LUKS containers in several different ways. Those
responsible at Ubuntu seem not to care very much (it is very easy to
recognize a LUKS container), so treat the process of installing Ubuntu
as a severe hazard to any LUKS container you may have.
NO WARNING ON NON-INTERACTIVE FORMAT: If you feed cryptsetup from STDIN
(e.g. via GnuPG) on LUKS format, it does not give you the warning that
you are about to format (and e.g. will lose any pre-existing LUKS
container on the target), as it assumes it is used from a script. In
this scenario, the responsibility for warning the user and possibly
checking for an existing LUKS header is shifted to the script. This is
a more general form of the previous item.
LUKS PASSPHRASE IS NOT THE MASTER KEY: The LUKS passphrase is not used
in deriving the master key. It is used in decrypting a master key that
is randomly selected on header creation. This means that if you create
a new LUKS header on top of an old one with exactly the same parameters
and exactly the same passphrase as the old one, it will still have a
different master key and your data will be permanently lost.
PASSPHRASE CHARACTER SET: Some people have had difficulties with this
when upgrading distributions. It is highly advisable to only use the 95
printable characters from the first 128 characters of the ASCII table,
as they will always have the same binary representation. Other
characters may have different encoding depending on system configuration
and your passphrase will not work with a different encoding. A table of
the standardized first 128 ASCII characters can, e.g. be found on
http://en.wikipedia.org/wiki/ASCII
KEYBOARD NUM-PAD: Apparently some pre-boot authentication environments
(these are done by the distro, not by cryptsetup, so complain there)
treat digits entered on the num-pad and ones entered regularly
different. This may be because the BIOS USB keyboard driver is used and
that one may have bugs on some computers. If you cannot open your
device in pre-boot, try entering the digits over the regular digit keys.
* 1.3 System specific warnings
- The Ubuntu Natty uinstaller has a "won't fix" defect that may destroy
LUKS containers. This is quite old an not relevant for most people.
Reference:
https://bugs.launchpad.net/ubuntu/+source/partman-crypto/+bug/420080
* 1.4 My LUKS-device is broken! Help!
First: Do not panic! In many cases the data is still recoverable.
Do not do anything hasty! Steps:
- Take some deep breaths. Maybe add some relaxing music. This may
sound funny, but I am completely serious. Often, critical damage is
done only after the initial problem.
- Do not reboot. The keys may still be in the kernel if the device is
mapped.
- Make sure others do not reboot the system.
- Do not write to your disk without a clear understanding why this will
not make matters worse. Do a sector-level backup before any writes.
Often you do not need to write at all to get enough access to make a
backup of the data.
- Relax some more.
- Read section 6 of this FAQ.
- Ask on the mailing-list if you need more help.
* 1.5 Who wrote this?
Current FAQ maintainer is Arno Wagner <arno@wagner.name>. If you want
to send me encrypted email, my current PGP key is DSA key CB5D9718,
fingerprint 12D6 C03B 1B30 33BB 13CF B774 E35C 5FA1 CB5D 9718.
Other contributors are listed at the end. If you want to contribute,
send your article, including a descriptive headline, to the maintainer,
or the dm-crypt mailing list with something like "FAQ ..."
in the subject. You can also send more raw information and have
me write the section. Please note that by contributing to this FAQ,
you accept the license described below.
This work is under the "Attribution-Share Alike 3.0 Unported" license,
which means distribution is unlimited, you may create derived works, but
attributions to original authors and this license statement must be
retained and the derived work must be under the same license. See
http://creativecommons.org/licenses/by-sa/3.0/ for more details of the
license.
Side note: I did text license research some time ago and I think this
license is best suited for the purpose at hand and creates the least
problems.
* 1.6 Where is the project website?
There is the project website at
https://gitlab.com/cryptsetup/cryptsetup/ Please do not post
questions there, nobody will read them. Use the mailing-list
instead.
* 1.7 Is there a mailing-list?
Instructions on how to subscribe to the mailing-list are at on the
project website. People are generally helpful and friendly on the
list.
The question of how to unsubscribe from the list does crop up sometimes.
For this you need your list management URL, which is sent to you
initially and once at the start of each month. Go to the URL mentioned
in the email and select "unsubscribe". This page also allows you to
request a password reminder.
Alternatively, you can send an Email to dm-crypt-request@saout.de with
just the word "help" in the subject or message body. Make sure to send
it from your list address.
The mailing list archive is here:
https://marc.info/?l=dm-crypt
* 1.8 Unsubscribe from the mailing-list
Send mail to dm-crypt-unsubscribe@saout.de from the subscribed account.
You will get an email with instructions.
Basically, you just have to respond to it unmodified to get
unsubscribed. The listserver admin functions are not very fast. It can
take 15 minutes or longer for a reply to arrive (I suspect greylisting
is in use), so be patient.
Also note that nobody on the list can unsubscribe you, sending demands
to be unsubscribed to the list just annoys people that are entirely
blameless for you being subscribed.
If you are subscribed, a subscription confirmation email was sent to
your email account and it had to be answered before the subscription
went active. The confirmation emails from the listserver have subjects
like these (with other numbers):
Subject: confirm 9964cf10.....
and are sent from dm-crypt-request@saout.de. You should check whether
you have anything like it in your sent email folder. If you find
nothing and are sure you did not confirm, then you should look into a
possible compromise of your email account.
2. Setup
* 2.1 LUKS Container Setup mini-HOWTO
This item tries to give you a very brief list of all the steps you
should go though when creating a new LUKS encrypted container, i.e.
encrypted disk, partition or loop-file.
01) All data will be lost, if there is data on the target, make a
backup.
02) Make very sure you use the right target disk, partition or
loop-file.
03) If the target was in use previously, it is a good idea to wipe it
before creating the LUKS container in order to remove any trace of old
file systems and data. For example, some users have managed to run
e2fsck on a partition containing a LUKS container, possibly because of
residual ext2 superblocks from an earlier use. This can do arbitrary
damage up to complete and permanent loss of all data in the LUKS
container.
To just quickly wipe file systems (old data may remain), use
wipefs -a <target device>
To wipe file system and data, use something like
cat /dev/zero > <target device>
This can take a while. To get a progress indicator, you can use the
tool dd_rescue (->google) instead or use my stream meter "wcs" (source
here: http://www.tansi.org/tools/index.html) in the following fashion:
cat /dev/zero | wcs > <target device>
Plain "dd" also gives you the progress on a SIGUSR1, see its man-page.
Be very sure you have the right target, all data will be lost!
Note that automatic wiping is on the TODO list for cryptsetup, so at
some time in the future this will become unnecessary.
Alternatively, plain dm-crypt can be used for a very fast wipe with
crypto-grade randomness, see Item 2.19
04) Create the LUKS container.
LUKS1:
cryptsetup luksFormat --type luks1 <target device>
LUKS2:
cryptsetup luksFormat --type luks2 <target device>
Just follow the on-screen instructions.
Note: Passprase iteration count is based on time and hence security
level depends on CPU power of the system the LUKS container is created
on. For example on a Raspberry Pi and LUKS1, I found some time ago that
the iteration count is 15 times lower than for a regular PC (well, for
my old one). Depending on security requirements, this may need
adjustment. For LUKS1, you can just look at the iteration count on
different systems and select one you like. You can also change the
benchmark time with the -i parameter to create a header for a slower
system.
For LUKS2, the parameters are more complex. ARGON2 has iteration,
parallelism and memory parameter. cryptsetup actually may adjust the
memory parameter for time scaling. Hence to use -i is the easiest way
to get slower or faster opening (default: 2000 = 2sec). Just make sure
to not drop this too low or you may get a memory parameter that is to
small to be secure. The luksDump command lists the memory parameter of
a created LUKS2 keyslot in kB. That parameter should probably be not
much lower than 100000, i.e. 100MB, but don't take my word for it.
05) Map the container. Here it will be mapped to /dev/mapper/c1:
cryptsetup luksOpen <target device> c1
06) (Optionally) wipe the container (make sure you have the right
target!):
cat /dev/zero > /dev/mapper/c1
This will take a while. Note that this creates a small information
leak, as an attacker can determine whether a 512 byte block is zero if
the attacker has access to the encrypted container multiple times.
Typically a competent attacker that has access multiple times can
install a passphrase sniffer anyways, so this leakage is not very
significant. For getting a progress indicator, see step 03.
07) Create a file system in the mapped container, for example an
ext3 file system (any other file system is possible):
mke2fs -j /dev/mapper/c1
08) Mount your encrypted file system, here on /mnt:
mount /dev/mapper/c1 /mnt
09) Make a LUKS header backup and plan for a container backup.
See Section 6 for details.
Done. You can now use the encrypted file system to store data. Be sure
to read though the rest of the FAQ, these are just the very basics. In
particular, there are a number of mistakes that are easy to make, but
will compromise your security.
* 2.2 LUKS on partitions or raw disks? What about RAID?
Also see Item 2.8.
This is a complicated question, and made more so by the availability of
RAID and LVM. I will try to give some scenarios and discuss advantages
and disadvantages. Note that I say LUKS for simplicity, but you can do
all the things described with plain dm-crypt as well. Also note that
your specific scenario may be so special that most or even all things I
say below do not apply.
Be aware that if you add LVM into the mix, things can get very
complicated. Same with RAID but less so. In particular, data recovery
can get exceedingly difficult. Only add LVM if you have a really good
reason and always remember KISS is what separates an engineer from an
amateur. Of course, if you really need the added complexity, KISS is
satisfied. But be very sure as there is a price to pay for it. In
engineering, complexity is always the enemy and needs to be fought
without mercy when encountered.
Also consider using RAID instead of LVM, as at least with the old
superblock format 0.90, the RAID superblock is in the place (end of
disk) where the risk of it damaging the LUKS header is smallest and you
can have your array assembled by the RAID controller (i.e. the kernel),
as it should be. Use partition type 0xfd for that. I recommend staying
away from superblock formats 1.0, 1.1 and 1.2 unless you really need
them.
Scenarios:
(1) Encrypted partition: Just make a partition to your liking, and put
LUKS on top of it and a filesystem into the LUKS container. This gives
you isolation of differently-tasked data areas, just as ordinary
partitioning does. You can have confidential data, non-confidential
data, data for some specific applications, user-homes, root, etc.
Advantages are simplicity as there is a 1:1 mapping between partitions
and filesystems, clear security functionality and the ability to
separate data into different, independent (!) containers.
Note that you cannot do this for encrypted root, that requires an
initrd. On the other hand, an initrd is about as vulnerable to a
competent attacker as a non-encrypted root, so there really is no
security advantage to doing it that way. An attacker that wants to
compromise your system will just compromise the initrd or the kernel
itself. The better way to deal with this is to make sure the root
partition does not store any critical data and to move that to
additional encrypted partitions. If you really are concerned your root
partition may be sabotaged by somebody with physical access (who would
however strangely not, say, sabotage your BIOS, keyboard, etc.), protect
it in some other way. The PC is just not set-up for a really secure
boot-chain (whatever some people may claim).
(2) Fully encrypted raw block device: For this, put LUKS on the raw
device (e.g. /dev/sdb) and put a filesystem into the LUKS container, no
partitioning whatsoever involved. This is very suitable for things like
external USB disks used for backups or offline data-storage.
(3) Encrypted RAID: Create your RAID from partitions and/or full
devices. Put LUKS on top of the RAID device, just if it were an
ordinary block device. Applications are just the same as above, but you
get redundancy. (Side note as many people seem to be unaware of it: You
can do RAID1 with an arbitrary number of components in Linux.) See also
Item 2.8.
(4) Now, some people advocate doing the encryption below the RAID layer.
That has several serious problems. One is that suddenly debugging RAID
issues becomes much harder. You cannot do automatic RAID assembly
anymore. You need to keep the encryption keys for the different RAID
components in sync or manage them somehow. The only possible advantage
is that things may run a little faster as more CPUs do the encryption,
but if speed is a priority over security and simplicity, you are doing
this wrong anyways. A good way to mitigate a speed issue is to get a
CPU that does hardware AES as most do today.
* 2.3 How do I set up encrypted swap?
As things that are confidential can end up in swap (keys, passphrases,
etc. are usually protected against being swapped to disk, but other
things may not be), it may be advisable to do something about the issue.
One option is to run without swap, which generally works well in a
desktop-context. It may cause problems in a server-setting or under
special circumstances. The solution to that is to encrypt swap with a
random key at boot-time.
NOTE: This is for Debian, and should work for Debian-derived
distributions. For others you may have to write your own startup script
or use other mechanisms.
01) Add the swap partition to /etc/crypttab. A line like the
following should do it:
swap /dev/<partition> /dev/urandom swap,noearly
Warning: While Debian refuses to overwrite partitions with a filesystem
or RAID signature on it, as your disk IDs may change (adding or removing
disks, failure of disk during boot, etc.), you may want to take
additional precautions. Yes, this means that your kernel device names
like sda, sdb, ... can change between reboots! This is not a concern
if you have only one disk. One possibility is to make sure the
partition number is not present on additional disks or also swap there.
Another is to encapsulate the swap partition (by making it a 1-partition
RAID1 or by using LVM), as that gets a persistent identifier.
Specifying it directly by UUID does not work, unfortunately, as the UUID
is part of the swap signature and that is not visible from the outside
due to the encryption and in addition changes on each reboot with this
setup.
Note: Use /dev/random if you are paranoid or in a potential low-entropy
situation (embedded system, etc.). This may cause the operation to take
a long time during boot however. If you are in a "no entropy"
situation, you cannot encrypt swap securely. In this situation you
should find some entropy, also because nothing else using crypto will be
secure, like ssh, ssl or GnuPG.
Note: The "noearly" option makes sure things like LVM, RAID, etc. are
running. As swap is non-critical for boot, it is fine to start it late.
02) Add the swap partition to /etc/fstab. A line like the following
should do it:
/dev/mapper/swap none swap sw 0 0
That is it. Reboot or start it manually to activate encrypted swap.
Manual start would look like this:
/etc/init.d/crypdisks start
swapon /dev/mapper/swap
* 2.4 What is the difference between "plain" and LUKS format?
First, unless you happen to understand the cryptographic background
well, you should use LUKS. It does protect the user from a lot of
common mistakes. Plain dm-crypt is for experts.
Plain format is just that: It has no metadata on disk, reads all
parameters from the commandline (or the defaults), derives a master-key
from the passphrase and then uses that to de-/encrypt the sectors of the
device, with a direct 1:1 mapping between encrypted and decrypted
sectors.
Primary advantage is high resilience to damage, as one damaged encrypted
sector results in exactly one damaged decrypted sector. Also, it is not
readily apparent that there even is encrypted data on the device, as an
overwrite with crypto-grade randomness (e.g. from
/dev/urandom) looks exactly the same on disk.
Side-note: That has limited value against the authorities. In civilized
countries, they cannot force you to give up a crypto-key anyways. In
quite a few countries around the world, they can force you to give up
the keys (using imprisonment or worse to pressure you, sometimes without
due process), and in the worst case, they only need a nebulous
"suspicion" about the presence of encrypted data. Sometimes this
applies to everybody, sometimes only when you are suspected of having
"illicit data" (definition subject to change) and sometimes specifically
when crossing a border. Note that this is going on in countries like
the US and the UK to different degrees and sometimes with courts
restricting what the authorities can actually demand.
My advice is to either be ready to give up the keys or to not have
encrypted data when traveling to those countries, especially when
crossing the borders. The latter also means not having any high-entropy
(random) data areas on your disk, unless you can explain them and
demonstrate that explanation. Hence doing a zero-wipe of all free
space, including unused space, may be a good idea.
Disadvantages are that you do not have all the nice features that the
LUKS metadata offers, like multiple passphrases that can be changed, the
cipher being stored in the metadata, anti-forensic properties like
key-slot diffusion and salts, etc..
LUKS format uses a metadata header and 8 key-slot areas that are being
placed at the beginning of the disk, see below under "What does the LUKS
on-disk format looks like?". The passphrases are used to decrypt a
single master key that is stored in the anti-forensic stripes. LUKS2
adds some more flexibility.
Advantages are a higher usability, automatic configuration of
non-default crypto parameters, defenses against low-entropy passphrases
like salting and iterated PBKDF2 or ARGON 2 passphrase hashing, the
ability to change passphrases, and others.
Disadvantages are that it is readily obvious there is encrypted data on
disk (but see side note above) and that damage to the header or
key-slots usually results in permanent data-loss. See below under "6.
Backup and Data Recovery" on how to reduce that risk. Also the sector
numbers get shifted by the length of the header and key-slots and there
is a loss of that size in capacity. Unless you have a specific need,
use LUKS2.
* 2.5 Can I encrypt an existing, non-empty partition to use LUKS?
There is no converter, and it is not really needed. The way to do this
is to make a backup of the device in question, securely wipe the device
(as LUKS device initialization does not clear away old data), do a
luksFormat, optionally overwrite the encrypted device, create a new
filesystem and restore your backup on the now encrypted device. Also
refer to sections "Security Aspects" and "Backup and Data Recovery".
For backup, plain GNU tar works well and backs up anything likely to be
in a filesystem.
* 2.6 How do I use LUKS with a loop-device?
This can be very handy for experiments. Setup is just the same as with
any block device. If you want, for example, to use a 100MiB file as
LUKS container, do something like this:
head -c 100M /dev/zero > luksfile # create empty file
losetup /dev/loop0 luksfile # map file to /dev/loop0
cryptsetup luksFormat --type luks2 /dev/loop0 # create LUKS2 container
Afterwards just use /dev/loop0 as a you would use a LUKS partition.
To unmap the file when done, use "losetup -d /dev/loop0".
* 2.7 When I add a new key-slot to LUKS, it asks for a passphrase
but then complains about there not being a key-slot with that
passphrase?
That is as intended. You are asked a passphrase of an existing key-slot
first, before you can enter the passphrase for the new key-slot.
Otherwise you could break the encryption by just adding a new key-slot.
This way, you have to know the passphrase of one of the already
configured key-slots in order to be able to configure a new key-slot.
* 2.8 Encryption on top of RAID or the other way round?
Also see Item 2.2.
Unless you have special needs, place encryption between RAID and
filesystem, i.e. encryption on top of RAID. You can do it the other
way round, but you have to be aware that you then need to give the
passphrase for each individual disk and RAID auto-detection will not
work anymore. Therefore it is better to encrypt the RAID device, e.g.
/dev/dm0 .
This means that the typical layering looks like this:
Filesystem <- top
|
Encryption (LUKS)
|
RAID
|
Raw partitions (optional)
|
Raw disks <- bottom
The big advantage of this is that you can manage the RAID container just
like any other regular RAID container, it does not care that its content
is encrypted. This strongly cuts down on complexity, something very
valuable with storage encryption.
* 2.9 How do I read a dm-crypt key from file?
Use the --key-file option, like this:
cryptsetup create --key-file keyfile e1 /dev/loop0
This will read the binary key from file, i.e. no hashing or
transformation will be applied to the keyfile before its bits are used
as key. Extra bits (beyond the length of the key) at the end are
ignored. Note that if you read from STDIN, the data will be hashed,
just as a key read interactively from the terminal. See the man-page
sections "NOTES ON PASSPHRASE PROCESSING..." for more detail.
* 2.10 How do I read a LUKS slot key from file?
What you really do here is to read a passphrase from file, just as you
would with manual entry of a passphrase for a key-slot. You can add a
new passphrase to a free key-slot, set the passphrase of an specific
key-slot or put an already configured passphrase into a file. Make sure
no trailing newline (0x0a) is contained in the input key file, or the
passphrase will not work because the whole file is used as input.
To add a new passphrase to a free key slot from file, use something
like this:
cryptsetup luksAddKey /dev/loop0 keyfile
To add a new passphrase to a specific key-slot, use something
like this:
cryptsetup luksAddKey --key-slot 7 /dev/loop0 keyfile
To supply a key from file to any LUKS command, use the --key-file
option, e.g. like this:
cryptsetup luksOpen --key-file keyfile /dev/loop0 e1
* 2.11 How do I read the LUKS master key from file?
The question you should ask yourself first is why you would want to do
this. The only legitimate reason I can think of is if you want to have
two LUKS devices with the same master key. Even then, I think it would
be preferable to just use key-slots with the same passphrase, or to use
plain dm-crypt instead. If you really have a good reason, please tell
me. If I am convinced, I will add how to do this here.
* 2.12 What are the security requirements for a key read from file?
A file-stored key or passphrase has the same security requirements as
one entered interactively, however you can use random bytes and thereby
use bytes you cannot type on the keyboard. You can use any file you
like as key file, for example a plain text file with a human readable
passphrase. To generate a file with random bytes, use something like
this:
head -c 256 /dev/random > keyfile
* 2.13 If I map a journaled file system using dm-crypt/LUKS, does
it still provide its usual transactional guarantees?
Yes, it does, unless a very old kernel is used. The required flags come
from the filesystem layer and are processed and passed onward by
dm-crypt (regardless of direct key management or LUKS key management).
A bit more information on the process by which transactional guarantees
are implemented can be found here:
http://lwn.net/Articles/400541/
Please note that these "guarantees" are weaker than they appear to be.
One problem is that quite a few disks lie to the OS about having flushed
their buffers. This is likely still true with SSDs. Some other things
can go wrong as well. The filesystem developers are aware of these
problems and typically can make it work anyways. That said,
dm-crypt/LUKS will not make things worse.
One specific problem you can run into is that you can get short freezes
and other slowdowns due to the encryption layer. Encryption takes time
and forced flushes will block for that time. For example, I did run
into frequent small freezes (1-2 sec) when putting a vmware image on
ext3 over dm-crypt. When I went back to ext2, the problem went away.
This seems to have gotten better with kernel 2.6.36 and the reworking of
filesystem flush locking mechanism (less blocking of CPU activity during
flushes). This should improve further and eventually the problem should
go away.
* 2.14 Can I use LUKS or cryptsetup with a more secure (external)
medium for key storage, e.g. TPM or a smartcard?
Yes, see the answers on using a file-supplied key. You do have to write
the glue-logic yourself though. Basically you can have cryptsetup read
the key from STDIN and write it there with your own tool that in turn
gets the key from the more secure key storage.
For TPM support, you may want to have a look at tpm-luks at
https://github.com/shpedoikal/tpm-luks. Note that tpm-luks is not
related to the cryptsetup project.
* 2.15 Can I resize a dm-crypt or LUKS container?
Yes, you can, as neither dm-crypt nor LUKS1 stores partition size and
LUKS2 uses a generic "whole device" size as default. Note that LUKS2
can use specified data-area sizes as a non-standard case and that these
may cause issues when resizing a LUKS2 container if set to a specific
value.
Whether you should do this is a different question. Personally I
recommend backup, recreation of the dm-crypt or LUKS container with new
size, recreation of the filesystem and restore. This gets around the
tricky business of resizing the filesystem. Resizing a dm-crypt or LUKS
container does not resize the filesystem in it. A backup is really
non-optional here, as a lot can go wrong, resulting in partial or
complete data loss. But if you have that backup, you can also just
recreate everything.
You also need to be aware of size-based limitations. The one currently
relevant is that aes-xts-plain should not be used for encrypted
container sizes larger than 2TiB. Use aes-xts-plain64 for that.
* 2.16 How do I Benchmark the Ciphers, Hashes and Modes?
Since version 1.60 cryptsetup supports the "benchmark" command.
Simply run as root:
cryptsetup benchmark
You can get more than the default benchmarks, see the man-page for the
relevant parameters. Note that XTS mode takes two keys, hence the
listed key sizes are double that for other modes and half of it is the
cipher key, the other half is the XTS key.
* 2.17 How do I Verify I have an Authentic cryptsetup Source Package?
Current maintainer is Milan Broz and he signs the release packages with
his PGP key. The key he currently uses is the "RSA key ID D93E98FC",
fingerprint 2A29 1824 3FDE 4664 8D06 86F9 D9B0 577B D93E 98FC. While I
have every confidence this really is his key and that he is who he
claims to be, don't depend on it if your life is at stake. For that
matter, if your life is at stake, don't depend on me being who I claim
to be either.
That said, as cryptsetup is under good version control and a malicious
change should be noticed sooner or later, but it may take a while.
Also, the attacker model makes compromising the sources in a non-obvious
way pretty hard. Sure, you could put the master-key somewhere on disk,
but that is rather obvious as soon as somebody looks as there would be
data in an empty LUKS container in a place it should not be. Doing this
in a more nefarious way, for example hiding the master-key in the salts,
would need a look at the sources to be discovered, but I think that
somebody would find that sooner or later as well.
That said, this discussion is really a lot more complicated and longer
as an FAQ can sustain. If in doubt, ask on the mailing list.
* 2.18 Is there a concern with 4k Sectors?
Not from dm-crypt itself. Encryption will be done in 512B blocks, but
if the partition and filesystem are aligned correctly and the filesystem
uses multiples of 4kiB as block size, the dm-crypt layer will just
process 8 x 512B = 4096B at a time with negligible overhead. LUKS does
place data at an offset, which is 2MiB per default and will not break
alignment. See also Item 6.12 of this FAQ for more details. Note that
if your partition or filesystem is misaligned, dm-crypt can make the
effect worse though. Also note that SSDs typically have much larger
blocks internally (e.g. 128kB or even larger).
* 2.19 How can I wipe a device with crypto-grade randomness?
The conventional recommendation if you want to do more than just a
zero-wipe is to use something like
cat /dev/urandom > <taget-device>
That used to very slow and painful at 10-20MB/s on a fast computer, but
newer kernels can give you > 200MB/s (depending on hardware). An
alternative is using cryptsetup and a plain dm-crypt device with a
random key, which is fast and on the same level of security. The
defaults are quite enough.
For device set-up, do the following:
cryptsetup open --type plain -d /dev/urandom /dev/<device> target
This maps the container as plain under /dev/mapper/target with a random
password. For the actual wipe you have several options. Basically, you
pipe zeroes into the opened container that then get encrypted. Simple
wipe without progress-indicator:
cat /dev/zero > /dev/mapper/to_be_wiped
Progress-indicator by dd_rescue:
dd_rescue -w /dev/zero /dev/mapper/to_be_wiped
Progress-indicator by my "wcs" stream meter (available from
http://www.tansi.org/tools/index.html ):
cat /dev/zero | wcs > /dev/mapper/to_be_wiped
Or use plain "dd", which gives you the progress when sent a SIGUSR1, see
the dd man page.
Remove the mapping at the end and you are done.
* 2.20 How to I wipe only the LUKS header?
This does _not_ describe an emergency wipe procedure, see Item 5.4 for
that. This procedure here is intended to be used when the data should
stay intact, e.g. when you change your LUKS container to use a detached
header and want to remove the old one. Please only do this if you have
a current backup.
LUKS1:
01) Determine header size in 512 Byte sectors with luksDump:
cryptsetup luksDump <device with LUKS container>
-> ...
Payload offset: <number>
...
02) Take the result number, multiply by 512 zeros and write to
the start of the device, e.g. like this:
dd bs=512 count=<number> if=/dev/zero of=<device>
LUKS2: (warning, untested! Remember that backup?) This assumes the
LUKS2 container uses the defaults, in particular there is only one data
segment. 01) Determine the data-segment offset using luksDump, same
as above for LUKS1:
-> ...
Data segments:
0: crypt
offset: <number> [bytes]
...
02) Overwrite the stated number of bytes from the start of the device.
Just to give yet another way to get a defined number of zeros:
head -c /dev/zero > /dev/<device>
3. Common Problems
* 3.1 My dm-crypt/LUKS mapping does not work! What general steps
are there to investigate the problem?
If you get a specific error message, investigate what it claims first.
If not, you may want to check the following things.
- Check that "/dev", including "/dev/mapper/control" is there. If it is
missing, you may have a problem with the "/dev" tree itself or you may
have broken udev rules.
- Check that you have the device mapper and the crypt target in your
kernel. The output of "dmsetup targets" should list a "crypt" target.
If it is not there or the command fails, add device mapper and
crypt-target to the kernel.
- Check that the hash-functions and ciphers you want to use are in the
kernel. The output of "cat /proc/crypto" needs to list them.
* 3.2 My dm-crypt mapping suddenly stopped when upgrading cryptsetup.
The default cipher, hash or mode may have changed (the mode changed from
1.0.x to 1.1.x). See under "Issues With Specific Versions of
cryptsetup".
* 3.3 When I call cryptsetup from cron/CGI, I get errors about
unknown features?
If you get errors about unknown parameters or the like that are not
present when cryptsetup is called from the shell, make sure you have no
older version of cryptsetup on your system that then gets called by
cron/CGI. For example some distributions install cryptsetup into
/usr/sbin, while a manual install could go to /usr/local/sbin. As a
debugging aid, call "cryptsetup --version" from cron/CGI or the
non-shell mechanism to be sure the right version gets called.
* 3.4 Unlocking a LUKS device takes very long. Why?
The unlock time for a key-slot (see Section 5 for an explanation what
iteration does) is calculated when setting a passphrase. By default it
is 1 second (2 seconds for LUKS2). If you set a passphrase on a fast
machine and then unlock it on a slow machine, the unlocking time can be
much longer. Also take into account that up to 8 key-slots (LUKS2: up
to 32 key-slots) have to be tried in order to find the right one.
If this is problem, you can add another key-slot using the slow machine
with the same passphrase and then remove the old key-slot. The new
key-slot will have the unlock time adjusted to the slow machine. Use
luksKeyAdd and then luksKillSlot or luksRemoveKey. You can also use
the -i option to reduce iteration time (and security level) when setting
a passphrase. Default is 1000 (1 sec) for LUKS1 and 2000 (2sec) for
LUKS2.
However, this operation will not change volume key iteration count ("MK
iterations" for LUKS1, "Iterations" under "Digests" for LUKS2). In
order to change that, you will have to backup the data in the LUKS
container (i.e. your encrypted data), luksFormat on the slow machine
and restore the data. Note that MK iterations are not very security
relevant.
* 3.5 "blkid" sees a LUKS UUID and an ext2/swap UUID on the same
device. What is wrong?
Some old versions of cryptsetup have a bug where the header does not get
completely wiped during LUKS format and an older ext2/swap signature
remains on the device. This confuses blkid.
Fix: Wipe the unused header areas by doing a backup and restore of
the header with cryptsetup 1.1.x or later:
cryptsetup luksHeaderBackup --header-backup-file <file> <device>
cryptsetup luksHeaderRestore --header-backup-file <file> <device>
4. Troubleshooting
* 4.1 I get the error "LUKS keyslot x is invalid." What does that mean?
For LUKS1, this means that the given keyslot has an offset that points
outside the valid keyslot area. Typically, the reason is a corrupted
LUKS1 header because something was written to the start of the device
the LUKS1 container is on. For LUKS2, I do not know when this error can
happen, but I expect it will be something similar. Refer to Section
"Backup and Data Recovery" and ask on the mailing list if you have
trouble diagnosing and (if still possible) repairing this.
* 4.2 I cannot unlock my LUKS container! What could be the problem?
First, make sure you have a correct passphrase. Then make sure you have
the correct key-map and correct keyboard. And then make sure you have
the correct character set and encoding, see also "PASSPHRASE CHARACTER
SET" under Section 1.2.
If you are sure you are entering the passphrase right, there is the
possibility that the respective key-slot has been damaged. There is no
way to recover a damaged key-slot, except from a header backup (see
Section 6). For security reasons, there is also no checksum in the
key-slots that could tell you whether a key-slot has been damaged. The
only checksum present allows recognition of a correct passphrase, but
that only works with that correct passphrase and a respective key-slot
that is intact.
In order to find out whether a key-slot is damaged one has to look for
"non-random looking" data in it. There is a tool that automatizes this
for LUKS1 in the cryptsetup distribution from version 1.6.0 onwards. It
is located in misc/keyslot_checker/. Instructions how to use and how to
interpret results are in the README file. Note that this tool requires
a libcryptsetup from cryptsetup 1.6.0 or later (which means
libcryptsetup.so.4.5.0 or later). If the tool complains about missing
functions in libcryptsetup, you likely have an earlier version from your
distribution still installed. You can either point the symbolic link(s)
from libcryptsetup.so.4 to the new version manually, or you can
uninstall the distribution version of cryptsetup and re-install that
from cryptsetup >= 1.6.0 again to fix this.
* 4.3 Can a bad RAM module cause problems?
LUKS and dm-crypt can give the RAM quite a workout, especially when
combined with software RAID. In particular the combination RAID5 +
LUKS1 + XFS seems to uncover RAM problems that do not cause obvious
problems otherwise. Symptoms vary, but often the problem manifest
itself when copying large amounts of data, typically several times
larger than your main memory.
Note: One thing you should always do on large data copying or movements
is to run a verify, for example with the "-d" option of "tar" or by
doing a set of MD5 checksums on the source or target with
find . -type f -exec md5sum \{\} \; > checksum-file
and then a "md5sum -c checksum-file" on the other side. If you get
mismatches here, RAM is the primary suspect. A lesser suspect is an
overclocked CPU. I have found countless hardware problems in verify
runs after copying data or making backups. Bit errors are much more
common than most people think.
Some RAM issues are even worse and corrupt structures in one of the
layers. This typically results in lockups, CPU state dumps in the
system logs, kernel panic or other things. It is quite possible to have
a problem with an encrypted device, but not with an otherwise the same
unencrypted device. The reason for that is that encryption has an error
amplification property: If you flip one bit in an encrypted data block,
the decrypted version has half of its bits flipped. This is actually an
important security property for modern ciphers. With the usual modes in
cryptsetup (CBC, ESSIV, XTS), you can get a completely changed 512 byte
block for a bit error. A corrupt block causes a lot more havoc than the
occasionally flipped single bit and can result in various obscure
errors.
Note that a verify run on copying between encrypted or unencrypted
devices will reliably detect corruption, even when the copying itself
did not report any problems. If you find defect RAM, assume all backups
and copied data to be suspect, unless you did a verify.
* 4.4 How do I test RAM?
First you should know that overclocking often makes memory problems
worse. So if you overclock (which I strongly recommend against in a
system holding data that has any worth), run the tests with the
overclocking active.
There are two good options. One is Memtest86+ and the other is
"memtester" by Charles Cazabon. Memtest86+ requires a reboot and then
takes over the machine, while memtester runs from a root-shell. Both
use different testing methods and I have found problems fast with either
one that the other needed long to find. I recommend running the
following procedure until the first error is found:
- Run Memtest86+ for one cycle
- Run memtester for one cycle (shut down as many other applications
as possible and use the largest memory area you can get)
- Run Memtest86+ for 24h or more
- Run memtester for 24h or more
If all that does not produce error messages, your RAM may be sound,
but I have had one weak bit in the past that Memtest86+ needed around
60 hours to find. If you can reproduce the original problem reliably,
a good additional test may be to remove half of the RAM (if you have
more than one module) and try whether the problem is still there and if
so, try with the other half. If you just have one module, get a
different one and try with that. If you do overclocking, reduce the
settings to the most conservative ones available and try with that.
* 4.5 Is there a risk using debugging tools like strace?
There most definitely is. A dump from strace and friends can contain
all data entered, including the full passphrase. Example with strace
and passphrase "test":
> strace cryptsetup luksOpen /dev/sda10 c1
...
read(6, "test\n", 512) = 5
...
Depending on different factors and the tool used, the passphrase may
also be encoded and not plainly visible. Hence it is never a good idea
to give such a trace from a live container to anybody. Recreate the
problem with a test container or set a temporary passphrase like "test"
and use that for the trace generation. Item 2.6 explains how to create
a loop-file backed LUKS container that may come in handy for this
purpose.
See also Item 6.10 for another set of data you should not give to
others.
5. Security Aspects
* 5.1 How long is a secure passphrase ?
This is just the short answer. For more info and explanation of some of
the terms used in this item, read the rest of Section 5. The actual
recommendation is at the end of this item.
First, passphrase length is not really the right measure, passphrase
entropy is. If your passphrase is 200 times the letter "a", it is long
but has very low entropy and is pretty insecure.
For example, a random lowercase letter (a-z) gives you 4.7 bit of
entropy, one element of a-z0-9 gives you 5.2 bits of entropy, an element
of a-zA-Z0-9 gives you 5.9 bits and a-zA-Z0-9!@#$%\^&:-+ gives you 6.2
bits. On the other hand, a random English word only gives you 0.6...1.3
bits of entropy per character. Using sentences that make sense gives
lower entropy, series of random words gives higher entropy. Do not use
sentences that can be tied to you or found on your computer. This type
of attack is done routinely today.
That said, it does not matter too much what scheme you use, but it does
matter how much entropy your passphrase contains, because an attacker
has to try on average
1/2 * 2^(bits of entropy in passphrase)
different passphrases to guess correctly.
Historically, estimations tended to use computing time estimates, but
more modern approaches try to estimate cost of guessing a passphrase.
As an example, I will try to get an estimate from the numbers in
https://gist.github.com/epixoip/a83d38f412b4737e99bbef804a270c40 This
thing costs 23kUSD and does 68Ghashes/sec for SHA1. This is in 2017.
Incidentally, my older calculation for a machine around 1000 times
slower was off by a factor of about 1000, but in the right direction,
i.e. I estimated the attack to be too easy. Nobody noticed ;-) On the
plus side, the tables are now (2017) pretty much accurate.
More references can be found a the end of this document. Note that
these are estimates from the defender side, so assuming something is
easier than it actually is is fine. An attacker may still have
significantly higher cost than estimated here.
LUKS1 used SHA1 (since version 1.7.0 it uses SHA256) for hashing per
default. We will leave aside the check whether a try actually decrypts
a key-slot. I will assume a useful lifetime of the hardware of 2 years.
(This is on the low side.) Disregarding downtime, the machine can then
break
N = 68*10^9 * 3600 * 24 * 365 * 2 ~ 4*10^18
passphrases for EUR/USD 23k. That is one 62 bit passphrase hashed once
with SHA1 for EUR/USD 23k. This can be parallelized, it can be done
faster than 2 years with several of these machines.
For LUKS2, things look a bit better, as the advantage of using graphics
cards is massively reduced. Using the recommendations below should
hence be fine for LUKS2 as well and give a better security margin.
For plain dm-crypt (no hash iteration) this is it. This gives (with
SHA1, plain dm-crypt default is ripemd160 which seems to be slightly
slower than SHA1):
Passphrase entropy Cost to break
60 bit EUR/USD 6k
65 bit EUR/USD 200K
70 bit EUR/USD 6M
75 bit EUR/USD 200M
80 bit EUR/USD 6B
85 bit EUR/USD 200B
... ...
For LUKS1, you have to take into account hash iteration in PBKDF2.
For a current CPU, there are about 100k iterations (as can be queried
with ''cryptsetup luksDump''.
The table above then becomes:
Passphrase entropy Cost to break
50 bit EUR/USD 600k
55 bit EUR/USD 20M
60 bit EUR/USD 600M
65 bit EUR/USD 20B
70 bit EUR/USD 600B
75 bit EUR/USD 20T
... ...
Recommendation:
To get reasonable security for the next 10 years, it is a good idea
to overestimate by a factor of at least 1000.
Then there is the question of how much the attacker is willing to spend.
That is up to your own security evaluation. For general use, I will
assume the attacker is willing to spend up to 1 million EUR/USD. Then
we get the following recommendations:
Plain dm-crypt: Use > 80 bit. That is e.g. 17 random chars from a-z
or a random English sentence of > 135 characters length.
LUKS1 and LUKS2: Use > 65 bit. That is e.g. 14 random chars from a-z
or a random English sentence of > 108 characters length.
If paranoid, add at least 20 bit. That is roughly four additional
characters for random passphrases and roughly 32 characters for a
random English sentence.
* 5.2 Is LUKS insecure? Everybody can see I have encrypted data!
In practice it does not really matter. In most civilized countries you
can just refuse to hand over the keys, no harm done. In some countries
they can force you to hand over the keys if they suspect encryption.
The suspicion is enough, they do not have to prove anything. This is
for practical reasons, as even the presence of a header (like the LUKS
header) is not enough to prove that you have any keys. It might have
been an experiment, for example. Or it was used as encrypted swap with
a key from /dev/random. So they make you prove you do not have
encrypted data. Of course, if true, that is impossible and hence the
whole idea is not compatible with fair laws. Note that in this context,
countries like the US or the UK are not civilized and do not have fair
laws.
This means that if you have a large set of random-looking data, they can
already lock you up. Hidden containers (encryption hidden within
encryption), as possible with Truecrypt, do not help either. They will
just assume the hidden container is there and unless you hand over the
key, you will stay locked up. Don't have a hidden container? Though
luck. Anybody could claim that.
Still, if you are concerned about the LUKS header, use plain dm-crypt
with a good passphrase. See also Section 2, "What is the difference
between "plain" and LUKS format?"
* 5.3 Should I initialize (overwrite) a new LUKS/dm-crypt partition?
If you just create a filesystem on it, most of the old data will still
be there. If the old data is sensitive, you should overwrite it before
encrypting. In any case, not initializing will leave the old data there
until the specific sector gets written. That may enable an attacker to
determine how much and where on the partition data was written. If you
think this is a risk, you can prevent this by overwriting the encrypted
device (here assumed to be named "e1") with zeros like this:
dd_rescue -w /dev/zero /dev/mapper/e1
or alternatively with one of the following more standard commands:
cat /dev/zero > /dev/mapper/e1
dd if=/dev/zero of=/dev/mapper/e1
* 5.4 How do I securely erase a LUKS container?
For LUKS, if you are in a desperate hurry, overwrite the LUKS header and
key-slot area. For LUKS1 and LUKS2, just be generous and overwrite the
first 100MB. A single overwrite with zeros should be enough. If you
anticipate being in a desperate hurry, prepare the command beforehand.
Example with /dev/sde1 as the LUKS partition and default parameters:
head -c 100000000 /dev/zero > /dev/sde1; sync
A LUKS header backup or full backup will still grant access to most or
all data, so make sure that an attacker does not have access to backups
or destroy them as well.
Also note that SSDs and also some HDDs (SMR and hybrid HDDs, for
example) may not actually overwrite the header and only do that an
unspecified and possibly very long time later. The only way to be sure
there is physical destruction. If the situation permits, do both
overwrite and physical destruction.
If you have time, overwrite the whole drive with a single pass of random
data. This is enough for most HDDs. For SSDs or FLASH (USB sticks) or
SMR or hybrid drives, you may want to overwrite the whole drive several
times to be sure data is not retained. This is possibly still insecure
as the respective technologies are not fully understood in this regard.
Still, due to the anti-forensic properties of the LUKS key-slots, a
single overwrite could be enough. If in doubt, use physical destruction
in addition. Here is a link to some current research results on erasing
SSDs and FLASH drives:
http://www.usenix.org/events/fast11/tech/full_papers/Wei.pdf
Keep in mind to also erase all backups.
Example for a random-overwrite erase of partition sde1 done with
dd_rescue:
dd_rescue -w /dev/urandom /dev/sde1
* 5.5 How do I securely erase a backup of a LUKS partition or header?
That depends on the medium it is stored on. For HDD and SSD, use
overwrite with random data. For an SSD, FLASH drive (USB stick) hybrid
HDD or SMR HDD, you may want to overwrite the complete drive several
times and use physical destruction in addition, see last item. For
re-writable CD/DVD, a single overwrite should be enough, due to the
anti-forensic properties of the LUKS keyslots. For write-once media,
use physical destruction. For low security requirements, just cut the
CD/DVD into several parts. For high security needs, shred or burn the
medium.
If your backup is on magnetic tape, I advise physical destruction by
shredding or burning, after (!) overwriting . The problem with magnetic
tape is that it has a higher dynamic range than HDDs and older data may
well be recoverable after overwrites. Also write-head alignment issues
can lead to data not actually being deleted during overwrites.
The best option is to actually encrypt the backup, for example with
PGP/GnuPG and then just destroy all copies of the encryption key if
needed. Best keep them on paper, as that has excellent durability and
secure destruction is easy, for example by burning and then crushing the
ashes to a fine powder. A blender and water also works nicely.
* 5.6 What about backup? Does it compromise security?
That depends. See item 6.7.
* 5.7 Why is all my data permanently gone if I overwrite the LUKS header?
Overwriting the LUKS header in part or in full is the most common reason
why access to LUKS containers is lost permanently. Overwriting can be
done in a number of fashions, like creating a new filesystem on the raw
LUKS partition, making the raw partition part of a raid array and just
writing to the raw partition.
The LUKS1 header contains a 256 bit "salt" per key-slot and without that
no decryption is possible. While the salts are not secret, they are
key-grade material and cannot be reconstructed. This is a
cryptographically strong "cannot". From observations on the cryptsetup
mailing-list, people typically go though the usual stages of grief
(Denial, Anger, Bargaining, Depression, Acceptance) when this happens to
them. Observed times vary between 1 day and 2 weeks to complete the
cycle. Seeking help on the mailing-list is fine. Even if we usually
cannot help with getting back your data, most people found the feedback
comforting.
If your header does not contain an intact key-slot salt, best go
directly to the last stage ("Acceptance") and think about what to do
now. There is one exception that I know of: If your LUKS1 container is
still open, then it may be possible to extract the master key from the
running system. See Item "How do I recover the master key from a mapped
LUKS1 container?" in Section "Backup and Data Recovery".
For LUKS2, things are both better and worse. First, the salts are in a
less vulnerable position now. But, on the other hand, the keys of a
mapped (open) container are now stored in the kernel key-store, and
while there probably is some way to get them out of there, I am not sure
how much effort that needs.
* 5.8 What is a "salt"?
A salt is a random key-grade value added to the passphrase before it is
processed. It is not kept secret. The reason for using salts is as
follows: If an attacker wants to crack the password for a single LUKS
container, then every possible passphrase has to be tried. Typically an
attacker will not try every binary value, but will try words and
sentences from a dictionary.
If an attacker wants to attack several LUKS containers with the same
dictionary, then a different approach makes sense: Compute the resulting
slot-key for each dictionary element and store it on disk. Then the
test for each entry is just the slow unlocking with the slot key (say
0.00001 sec) instead of calculating the slot-key first (1 sec). For a
single attack, this does not help. But if you have more than one
container to attack, this helps tremendously, also because you can
prepare your table before you even have the container to attack! The
calculation is also very simple to parallelize. You could, for example,
use the night-time unused CPU power of your desktop PCs for this.
This is where the salt comes in. If the salt is combined with the
passphrase (in the simplest form, just appended to it), you suddenly
need a separate table for each salt value. With a reasonably-sized salt
value (256 bit, e.g.) this is quite infeasible.
* 5.9 Is LUKS secure with a low-entropy (bad) passphrase?
Short answer: yes. Do not use a low-entropy passphrase.
Note: For LUKS2, protection for bad passphrases is a bit better
due to the use of Argon2, but that is only a gradual improvement.
Longer answer:
This needs a bit of theory. The quality of your passphrase is directly
related to its entropy (information theoretic, not thermodynamic). The
entropy says how many bits of "uncertainty" or "randomness" are in you
passphrase. In other words, that is how difficult guessing the
passphrase is.
Example: A random English sentence has about 1 bit of entropy per
character. A random lowercase (or uppercase) character has about 4.7
bit of entropy.
Now, if n is the number of bits of entropy in your passphrase and t
is the time it takes to process a passphrase in order to open the
LUKS container, then an attacker has to spend at maximum
attack_time_max = 2^n * t
time for a successful attack and on average half that. There is no way
getting around that relationship. However, there is one thing that does
help, namely increasing t, the time it takes to use a passphrase, see
next FAQ item.
Still, if you want good security, a high-entropy passphrase is the only
option. For example, a low-entropy passphrase can never be considered
secure against a TLA-level (Three Letter Agency level, i.e.
government-level) attacker, no matter what tricks are used in the
key-derivation function. Use at least 64 bits for secret stuff. That
is 64 characters of English text (but only if randomly chosen) or a
combination of 12 truly random letters and digits.
For passphrase generation, do not use lines from very well-known texts
(religious texts, Harry potter, etc.) as they are too easy to guess.
For example, the total Harry Potter has about 1'500'000 words (my
estimation). Trying every 64 character sequence starting and ending at
a word boundary would take only something like 20 days on a single CPU
and is entirely feasible. To put that into perspective, using a number
of Amazon EC2 High-CPU Extra Large instances (each gives about 8 real
cores), this test costs currently about 50USD/EUR, but can be made to
run arbitrarily fast.
On the other hand, choosing 1.5 lines from, say, the Wheel of Time, is
in itself not more secure, but the book selection adds quite a bit of
entropy. (Now that I have mentioned it here, don't use tWoT either!) If
you add 2 or 3 typos and switch some words around, then this is good
passphrase material.
* 5.10 What is "iteration count" and why is decreasing it a bad idea?
LUKS1:
Iteration count is the number of PBKDF2 iterations a passphrase is put
through before it is used to unlock a key-slot. Iterations are done
with the explicit purpose to increase the time that it takes to unlock a
key-slot. This provides some protection against use of low-entropy
passphrases.
The idea is that an attacker has to try all possible passphrases. Even
if the attacker knows the passphrase is low-entropy (see last item), it
is possible to make each individual try take longer. The way to do this
is to repeatedly hash the passphrase for a certain time. The attacker
then has to spend the same time (given the same computing power) as the
user per try. With LUKS1, the default is 1 second of PBKDF2 hashing.
Example 1: Lets assume we have a really bad passphrase (e.g. a
girlfriends name) with 10 bits of entropy. With the same CPU, an
attacker would need to spend around 500 seconds on average to break that
passphrase. Without iteration, it would be more like 0.0001 seconds on
a modern CPU.
Example 2: The user did a bit better and has 32 chars of English text.
That would be about 32 bits of entropy. With 1 second iteration, that
means an attacker on the same CPU needs around 136 years. That is
pretty impressive for such a weak passphrase. Without the iterations,
it would be more like 50 days on a modern CPU, and possibly far less.
In addition, the attacker can both parallelize and use special hardware
like GPUs or FPGAs to speed up the attack. The attack can also happen
quite some time after the luksFormat operation and CPUs can have become
faster and cheaper. For that reason you want a bit of extra security.
Anyways, in Example 1 your are screwed. In example 2, not necessarily.
Even if the attack is faster, it still has a certain cost associated
with it, say 10000 EUR/USD with iteration and 1 EUR/USD without
iteration. The first can be prohibitively expensive, while the second
is something you try even without solid proof that the decryption will
yield something useful.
The numbers above are mostly made up, but show the idea. Of course the
best thing is to have a high-entropy passphrase.
Would a 100 sec iteration time be even better? Yes and no.
Cryptographically it would be a lot better, namely 100 times better.
However, usability is a very important factor for security technology
and one that gets overlooked surprisingly often. For LUKS, if you have
to wait 2 minutes to unlock the LUKS container, most people will not
bother and use less secure storage instead. It is better to have less
protection against low-entropy passphrases and people actually use LUKS,
than having them do without encryption altogether.
Now, what about decreasing the iteration time? This is generally a very
bad idea, unless you know and can enforce that the users only use
high-entropy passphrases. If you decrease the iteration time without
ensuring that, then you put your users at increased risk, and
considering how rarely LUKS containers are unlocked in a typical
work-flow, you do so without a good reason. Don't do it. The iteration
time is already low enough that users with low entropy passphrases are
vulnerable. Lowering it even further increases this danger
significantly.
LUKS2: Pretty much the same reasoning applies. The advantages of using
GPUs or FPGAs in an attack have been significantly reduced, but that
is the only main difference.
* 5.11 Some people say PBKDF2 is insecure?
There is some discussion that a hash-function should have a "large
memory" property, i.e. that it should require a lot of memory to be
computed. This serves to prevent attacks using special programmable
circuits, like FPGAs, and attacks using graphics cards. PBKDF2 does not
need a lot of memory and is vulnerable to these attacks. However, the
publication usually referred in these discussions is not very convincing
in proving that the presented hash really is "large memory" (that may
change, email the FAQ maintainer when it does) and it is of limited
usefulness anyways. Attackers that use clusters of normal PCs will not
be affected at all by a "large memory" property. For example the US
Secret Service is known to use the off-hour time of all the office PCs
of the Treasury for password breaking. The Treasury has about 110'000
employees. Assuming every one has an office PC, that is significant
computing power, all of it with plenty of memory for computing "large
memory" hashes. Bot-net operators also have all the memory they want.
The only protection against a resourceful attacker is a high-entropy
passphrase, see items 5.9 and 5.10.
That said, LUKS2 defaults to Argon2, which has a large-memory property
and massively reduces the advantages of GPUs and FPGAs.
* 5.12 What about iteration count with plain dm-crypt?
Simple: There is none. There is also no salting. If you use plain
dm-crypt, the only way to be secure is to use a high entropy passphrase.
If in doubt, use LUKS instead.
* 5.13 Is LUKS with default parameters less secure on a slow CPU?
Unfortunately, yes. However the only aspect affected is the protection
for low-entropy passphrase or master-key. All other security aspects
are independent of CPU speed.
The master key is less critical, as you really have to work at it to
give it low entropy. One possibility to mess this up is to supply the
master key yourself. If that key is low-entropy, then you get what you
deserve. The other known possibility to create a LUKS container with a
bad master key is to use /dev/urandom for key generation in an
entropy-starved situation (e.g. automatic installation on an embedded
device without network and other entropy sources or installation in a VM
under certain circumstances).
For the passphrase, don't use a low-entropy passphrase. If your
passphrase is good, then a slow CPU will not matter. If you insist on a
low-entropy passphrase on a slow CPU, use something like
"--iter-time=10000" or higher and wait a long time on each LUKS unlock
and pray that the attacker does not find out in which way exactly your
passphrase is low entropy. This also applies to low-entropy passphrases
on fast CPUs. Technology can do only so much to compensate for problems
in front of the keyboard.
Also note that power-saving modes will make your CPU slower. This will
reduce iteration count on LUKS container creation. It will keep unlock
times at the expected values though at this CPU speed.
* 5.14 Why was the default aes-cbc-plain replaced with aes-cbc-essiv?
Note: This item applies both to plain dm-crypt and to LUKS
The problem is that cbc-plain has a fingerprint vulnerability, where a
specially crafted file placed into the crypto-container can be
recognized from the outside. The issue here is that for cbc-plain the
initialization vector (IV) is the sector number. The IV gets XORed to
the first data chunk of the sector to be encrypted. If you make sure
that the first data block to be stored in a sector contains the sector
number as well, the first data block to be encrypted is all zeros and
always encrypted to the same ciphertext. This also works if the first
data chunk just has a constant XOR with the sector number. By having
several shifted patterns you can take care of the case of a
non-power-of-two start sector number of the file.
This mechanism allows you to create a pattern of sectors that have the
same first ciphertext block and signal one bit per sector to the
outside, allowing you to e.g. mark media files that way for recognition
without decryption. For large files this is a practical attack. For
small ones, you do not have enough blocks to signal and take care of
different file starting offsets.
In order to prevent this attack, the default was changed to cbc-essiv.
ESSIV uses a keyed hash of the sector number, with the encryption key as
key. This makes the IV unpredictable without knowing the encryption key
and the watermarking attack fails.
* 5.15 Are there any problems with "plain" IV? What is "plain64"?
First, "plain" and "plain64" are both not secure to use with CBC, see
previous FAQ item.
However there are modes, like XTS, that are secure with "plain" IV. The
next limit is that "plain" is 64 bit, with the upper 32 bit set to zero.
This means that on volumes larger than 2TiB, the IV repeats, creating a
vulnerability that potentially leaks some data. To avoid this, use
"plain64", which uses the full sector number up to 64 bit. Note that
"plain64" requires a kernel 2.6.33 or more recent. Also note that
"plain64" is backwards compatible for volume sizes of maximum size 2TiB,
but not for those > 2TiB. Finally, "plain64" does not cause any
performance penalty compared to "plain".
* 5.16 What about XTS mode?
XTS mode is potentially even more secure than cbc-essiv (but only if
cbc-essiv is insecure in your scenario). It is a NIST standard and
used, e.g. in Truecrypt. From version 1.6.0 of cryptsetup onwards,
aes-xts-plain64 is the default for LUKS. If you want to use it with a
cryptsetup before version 1.6.0 or with plain dm-crypt, you have to
specify it manually as "aes-xts-plain", i.e.
cryptsetup -c aes-xts-plain luksFormat <device>
For volumes >2TiB and kernels >= 2.6.33 use "plain64" (see FAQ item
on "plain" and "plain64"):
cryptsetup -c aes-xts-plain64 luksFormat <device>
There is a potential security issue with XTS mode and large blocks.
LUKS and dm-crypt always use 512B blocks and the issue does not apply.
* 5.17 Is LUKS FIPS-140-2 certified?
No. But that is more a problem of FIPS-140-2 than of LUKS. From a
technical point-of-view, LUKS with the right parameters would be
FIPS-140-2 compliant, but in order to make it certified, somebody has to
pay real money for that. And then, whenever cryptsetup is changed or
extended, the certification lapses and has to be obtained again.
From the aspect of actual security, LUKS with default parameters should
be as good as most things that are FIPS-140-2 certified, although you
may want to make sure to use /dev/random (by specifying --use-random on
luksFormat) as randomness source for the master key to avoid being
potentially insecure in an entropy-starved situation.
* 5.18 What about Plausible Deniability?
First let me attempt a definition for the case of encrypted filesystems:
Plausible deniability is when you store data inside an encrypted
container and it is not possible to prove it is there without having a
special passphrase. And at the same time it must be "plausible" that
there actually is no hidden data there.
As a simple entropy-analysis will show that here may be data there, the
second part is what makes it tricky.
There seem to be a lot of misunderstandings about this idea, so let me
make it clear that this refers to the situation where the attackers can
prove that there is data that either may be random or may be part of a
plausible-deniability scheme, they just cannot prove which one it is.
Hence a plausible-deniability scheme must hold up when the attackers
know there is something potentially fishy. If you just hide data and
rely on it not being found, that is just simple deniability, not
"plausible" deniability and I am not talking about that in the
following. Simple deniability against a low-competence attacker may be
as simple as renaming a file or putting data into an unused part of a
disk. Simple deniability against a high-skill attacker with time to
invest is usually pointless unless you go for advanced steganographic
techniques, which have their own drawbacks, such as low data capacity.
Now, the idea of plausible deniability is compelling and on a first
glance it seems possible to do it. And from a cryptographic point of
view, it actually is possible.
So, does the idea work in practice? No, unfortunately. The reasoning
used by its proponents is fundamentally flawed in several ways and the
cryptographic properties fail fatally when colliding with the real
world.
First, why should "I do not have a hidden partition" be any more
plausible than "I forgot my crypto key" or "I wiped that partition with
random data, nothing in there"? I do not see any reason.
Second, there are two types of situations: Either they cannot force you
to give them the key (then you simply do not) or they can. In the
second case, they can always do bad things to you, because they cannot
prove that you have the key in the first place! This means they do not
have to prove you have the key, or that this random looking data on your
disk is actually encrypted data. So the situation will allow them to
waterboard/lock-up/deport you anyways, regardless of how "plausible"
your deniability is. Do not have a hidden partition you could show to
them, but there are indications you may? Too bad for you.
Unfortunately "plausible deniability" also means you cannot prove there
is no hidden data.
Third, hidden partitions are not that hidden. There are basically just
two possibilities: a) Make a large crypto container, but put a smaller
filesystem in there and put the hidden partition into the free space.
Unfortunately this is glaringly obvious and can be detected in an
automated fashion. This means that the initial suspicion to put you
under duress in order to make you reveal your hidden data is given. b)
Make a filesystem that spans the whole encrypted partition, and put the
hidden partition into space not currently used by that filesystem.
Unfortunately that is also glaringly obvious, as you then cannot write
to the filesystem without a high risk of destroying data in the hidden
container. Have not written anything to the encrypted filesystem in a
while? Too bad, they have the suspicion they need to do unpleasant
things to you.
To be fair, if you prepare option b) carefully and directly before going
into danger, it may work. But then, the mere presence of encrypted data
may already be enough to get you into trouble in those places were they
can demand encryption keys.
Here is an additional reference for some problems with plausible
deniability: http://www.schneier.com/paper-truecrypt-dfs.pdf I strongly
suggest you read it.
So, no, I will not provide any instructions on how to do it with plain
dm-crypt or LUKS. If you insist on shooting yourself in the foot, you
can figure out how to do it yourself.
* 5.19 What about SSDs, Flash, Hybrid and SMR Drives?
The problem is that you cannot reliably erase parts of these devices,
mainly due to wear-leveling and possibly defect management and delayed
writes to the main data area.
For example for SSDs, when overwriting a sector, what the device does is
to move an internal sector (may be 128kB or even larger) to some pool of
discarded, not-yet erased unused sectors, take a fresh empty sector from
the empty-sector pool and copy the old sector over with the changes to
the small part you wrote. This is done in some fashion so that larger
writes do not cause a lot of small internal updates.
The thing is that the mappings between outside-addressable sectors and
inside sectors is arbitrary (and the vendors are not talking). Also the
discarded sectors are not necessarily erased immediately. They may
linger a long time.
For plain dm-crypt, the consequences are that older encrypted data may
be lying around in some internal pools of the device. Thus may or may
not be a problem and depends on the application. Remember the same can
happen with a filesystem if consecutive writes to the same area of a
file can go to different sectors.
However, for LUKS, the worst case is that key-slots and LUKS header may
end up in these internal pools. This means that password management
functionality is compromised (the old passwords may still be around,
potentially for a very long time) and that fast erase by overwriting the
header and key-slot area is insecure.
Also keep in mind that the discarded/used pool may be large. For
example, a 240GB SSD has about 16GB of spare area in the chips that it
is free to do with as it likes. You would need to make each individual
key-slot larger than that to allow reliable overwriting. And that
assumes the disk thinks all other space is in use. Reading the internal
pools using forensic tools is not that hard, but may involve some
soldering.
What to do?
If you trust the device vendor (you probably should not...) you can try
an ATA "secure erase" command. That is not present in USB keys though
and may or may not be secure for a hybrid drive.
If you can do without password management and are fine with doing
physical destruction for permanently deleting data (always after one or
several full overwrites!), you can use plain dm-crypt.
If you want or need all the original LUKS security features to work, you
can use a detached LUKS header and put that on a conventional, magnetic
disk. That leaves potentially old encrypted data in the pools on the
main disk, but otherwise you get LUKS with the same security as on a
traditional magnetic disk. Note however that storage vendors are prone
to lying to their customers. For example, it recently came out that
HDDs sold without any warning or mentioning in the data-sheets were
actually using SMR and that will write data first to a faster area and
only overwrite the original data area some time later when things are
quiet.
If you are concerned about your laptop being stolen, you are likely fine
using LUKS on an SSD or hybrid drive. An attacker would need to have
access to an old passphrase (and the key-slot for this old passphrase
would actually need to still be somewhere in the SSD) for your data to
be at risk. So unless you pasted your old passphrase all over the
Internet or the attacker has knowledge of it from some other source and
does a targeted laptop theft to get at your data, you should be fine.
* 5.20 LUKS1 is broken! It uses SHA-1!
No, it is not. SHA-1 is (academically) broken for finding collisions,
but not for using it in a key-derivation function. And that collision
vulnerability is for non-iterated use only. And you need the hash-value
in verbatim.
This basically means that if you already have a slot-key, and you have
set the PBKDF2 iteration count to 1 (it is > 10'000 normally), you could
(maybe) derive a different passphrase that gives you the the same
slot-key. But if you have the slot-key, you can already unlock the
key-slot and get the master key, breaking everything. So basically,
this SHA-1 vulnerability allows you to open a LUKS1 container with high
effort when you already have it open.
The real problem here is people that do not understand crypto and claim
things are broken just because some mechanism is used that has been
broken for a specific different use. The way the mechanism is used
matters very much. A hash that is broken for one use can be completely
secure for other uses and here it is.
Since version 1.7.0, cryptsetup uses SHA-256 as default to ensure that
it will be compatible in the future. There are already some systems
where SHA-1 is completely phased out or disabled by a security policy.
* 5.21 Why is there no "Nuke-Option"?
A "Nuke-Option" or "Kill-switch" is a password that when entered upon
unlocking instead wipes the header and all passwords. So when somebody
forces you to enter your password, you can destroy the data instead.
While this sounds attractive at first glance, it does not make sense
once a real security analysis is done. One problem is that you have to
have some kind of HSM (Hardware Security Module) in order to implement
it securely. In the movies, a HSM starts to smoke and melt once the
Nuke-Option has been activated. In actual reality, it just wipes some
battery-backed RAM cells. A proper HSM costs something like
20'000...100'000 EUR/USD and there a Nuke-Option may make some sense.
BTW, a chipcard or a TPM is not a HSM, although some vendors are
promoting that myth.
Now, a proper HSMs will have a wipe option but not a Nuke-Option, i.e.
you can explicitly wipe the HSM, but by a different process than
unlocking it takes. Why is that? Simple: If somebody can force you to
reveal passwords, then they can also do bad things to you if you do not
or if you enter a nuke password instead. Think locking you up for a few
years for "destroying evidence" or for far longer and without trial for
being a "terrorist suspect". No HSM maker will want to expose its
customers to that risk.
Now think of the typical LUKS application scenario, i.e. disk
encryption. Usually the ones forcing you to hand over your password
will have access to the disk as well, and, if they have any real
suspicion, they will mirror your disk before entering anything supplied
by you. This neatly negates any Nuke-Option. If they have no suspicion
(just harassing people that cross some border for example), the
Nuke-Option would work, but see above about likely negative consequences
and remember that a Nuke-Option may not work reliably on SSD and hybrid
drives anyways.
Hence my advice is to never take data that you do not want to reveal
into any such situation in the first place. There is no need to
transfer data on physical carriers today. The Internet makes it quite
possible to transfer data between arbitrary places and modern encryption
makes it secure. If you do it right, nobody will even be able to
identify source or destination. (How to do that is out of scope of this
document. It does require advanced skills in this age of pervasive
surveillance.)
Hence, LUKS has not kill option because it would do much more harm than
good.
Still, if you have a good use-case (i.e. non-abstract real-world
situation) where a Nuke-Option would actually be beneficial, please let
me know.
* 5.22 Does cryptsetup open network connections to websites, etc. ?
This question seems not to make much sense at first glance, but here is
an example form the real world: The TrueCrypt GUI has a "Donation"
button. Press it, and a web-connection to the TrueCrypt website is
opened via the default browser, telling everybody that listens that you
use TrueCrypt. In the worst case, things like this can get people
tortured or killed.
So: Cryptsetup will never open any network connections except the
local netlink socket it needs to talk to the kernel crypto API.
In addition, the installation package should contain all documentation,
including this FAQ, so that you do not have to go to a web-site to read
it. (If your distro cuts the docu, please complain to them.) In
security software, any connection initiated to anywhere outside your
machine should always be the result of an explicit request for such a
connection by the user and cryptsetup will stay true to that principle.
6. Backup and Data Recovery
* 6.1 Why do I need Backup?
First, disks die. The rate for well-treated (!) disk is about 5% per
year, which is high enough to worry about. There is some indication
that this may be even worse for some SSDs. This applies both to LUKS
and plain dm-crypt partitions.
Second, for LUKS, if anything damages the LUKS header or the key-stripe
area then decrypting the LUKS device can become impossible. This is a
frequent occurrence. For example an accidental format as FAT or some
software overwriting the first sector where it suspects a partition boot
sector typically makes a LUKS1 partition permanently inaccessible. See
more below on LUKS header damage.
So, data-backup in some form is non-optional. For LUKS, you may also
want to store a header backup in some secure location. This only needs
an update if you change passphrases.
* 6.2 How do I backup a LUKS header?
While you could just copy the appropriate number of bytes from the start
of the LUKS partition, the best way is to use command option
"luksHeaderBackup" of cryptsetup. This protects also against errors
when non-standard parameters have been used in LUKS partition creation.
Example:
cryptsetup luksHeaderBackup --header-backup-file <file> <device>
To restore, use the inverse command, i.e.
cryptsetup luksHeaderRestore --header-backup-file <file> <device>
If you are unsure about a header to be restored, make a backup of the
current one first! You can also test the header-file without restoring
it by using the --header option for a detached header like this:
cryptsetup --header <file> luksOpen <device> </dev/mapper/name>
If that unlocks your keys-lot, you are good. Do not forget to close
the device again.
Under some circumstances (damaged header), this fails. Then use the
following steps in case it is LUKS1:
First determine the master-key size:
cryptsetup luksDump <device>
gives a line of the form
MK bits: <bits>
with bits equal to 256 for the old defaults and 512 for the new
defaults. 256 bits equals a total header size of 1'052'672 Bytes and
512 bits one of 2MiB. (See also Item 6.12) If luksDump fails, assume
2MiB, but be aware that if you restore that, you may also restore the
first 1M or so of the filesystem. Do not change the filesystem if you
were unable to determine the header size! With that, restoring a
too-large header backup is still safe.
Second, dump the header to file. There are many ways to do it, I
prefer the following:
head -c 1052672 <device> > header_backup.dmp
or
head -c 2M <device> > header_backup.dmp
for a 2MiB header. Verify the size of the dump-file to be sure.
To restore such a backup, you can try luksHeaderRestore or do a more
basic
cat header_backup.dmp > <device>
* 6.3 How do I test for a LUKS header?
Use
cryptsetup -v isLuks <device>
on the device. Without the "-v" it just signals its result via
exit-status. You can also use the more general test
blkid -p <device>
which will also detect other types and give some more info. Omit
"-p" for old versions of blkid that do not support it.
* 6.4 How do I backup a LUKS or dm-crypt partition?
There are two options, a sector-image and a plain file or filesystem
backup of the contents of the partition. The sector image is already
encrypted, but cannot be compressed and contains all empty space. The
filesystem backup can be compressed, can contain only part of the
encrypted device, but needs to be encrypted separately if so desired.
A sector-image will contain the whole partition in encrypted form, for
LUKS the LUKS header, the keys-slots and the data area. It can be done
under Linux e.g. with dd_rescue (for a direct image copy) and with
"cat" or "dd". Examples:
cat /dev/sda10 > sda10.img
dd_rescue /dev/sda10 sda10.img
You can also use any other backup software that is capable of making a
sector image of a partition. Note that compression is ineffective for
encrypted data, hence it does not make sense to use it.
For a filesystem backup, you decrypt and mount the encrypted partition
and back it up as you would a normal filesystem. In this case the
backup is not encrypted, unless your encryption method does that. For
example you can encrypt a backup with "tar" as follows with GnuPG:
tar cjf - <path> | gpg --cipher-algo AES -c - > backup.tbz2.gpg
And verify the backup like this if you are at "path":
cat backup.tbz2.gpg | gpg - | tar djf -
Note: Always verify backups, especially encrypted ones!
There is one problem with verifying like this: The kernel may still have
some files cached and in fact verify them against RAM or may even verify
RAM against RAM, which defeats the purpose of the exercise. The
following command empties the kernel caches:
echo 3 > /proc/sys/vm/drop_caches
Run it after backup and before verify.
In both cases GnuPG will ask you interactively for your symmetric key.
The verify will only output errors. Use "tar dvjf -" to get all
comparison results. To make sure no data is written to disk
unencrypted, turn off swap if it is not encrypted before doing the
backup.
Restore works like certification with the 'd' ('difference') replaced
by 'x' ('eXtract'). Refer to the man-page of tar for more explanations
and instructions. Note that with default options tar will overwrite
already existing files without warning. If you are unsure about how
to use tar, experiment with it in a location where you cannot do damage.
You can of course use different or no compression and you can use an
asymmetric key if you have one and have a backup of the secret key that
belongs to it.
A second option for a filesystem-level backup that can be used when the
backup is also on local disk (e.g. an external USB drive) is to use a
LUKS container there and copy the files to be backed up between both
mounted containers. Also see next item.
* 6.5 Do I need a backup of the full partition? Would the header
and key-slots not be enough?
Backup protects you against two things: Disk loss or corruption and user
error. By far the most questions on the dm-crypt mailing list about how
to recover a damaged LUKS partition are related to user error. For
example, if you create a new filesystem on a non-mapped LUKS container,
chances are good that all data is lost permanently.
For this case, a header+key-slot backup would often be enough. But keep
in mind that a well-treated (!) HDD has roughly a failure risk of 5% per
year. It is highly advisable to have a complete backup to protect
against this case.
* 6.6 What do I need to backup if I use "decrypt_derived"?
This is a script in Debian, intended for mounting /tmp or swap with a
key derived from the master key of an already decrypted device. If you
use this for an device with data that should be persistent, you need to
make sure you either do not lose access to that master key or have a
backup of the data. If you derive from a LUKS device, a header backup
of that device would cover backing up the master key. Keep in mind that
this does not protect against disk loss.
Note: If you recreate the LUKS header of the device you derive from
(using luksFormat), the master key changes even if you use the same
passphrase(s) and you will not be able to decrypt the derived device
with the new LUKS header.
* 6.7 Does a backup compromise security?
Depends on how you do it. However if you do not have one, you are going
to eventually lose your encrypted data.
There are risks introduced by backups. For example if you
change/disable a key-slot in LUKS, a binary backup of the partition will
still have the old key-slot. To deal with this, you have to be able to
change the key-slot on the backup as well, securely erase the backup or
do a filesystem-level backup instead of a binary one.
If you use dm-crypt, backup is simpler: As there is no key management,
the main risk is that you cannot wipe the backup when wiping the
original. However wiping the original for dm-crypt should consist of
forgetting the passphrase and that you can do without actual access to
the backup.
In both cases, there is an additional (usually small) risk with binary
backups: An attacker can see how many sectors and which ones have been
changed since the backup. To prevent this, use a filesystem level
backup method that encrypts the whole backup in one go, e.g. as
described above with tar and GnuPG.
My personal advice is to use one USB disk (low value data) or three
disks (high value data) in rotating order for backups, and either use
independent LUKS partitions on them, or use encrypted backup with tar
and GnuPG.
If you do network-backup or tape-backup, I strongly recommend to go
the filesystem backup path with independent encryption, as you
typically cannot reliably delete data in these scenarios, especially
in a cloud setting. (Well, you can burn the tape if it is under your
control...)
* 6.8 What happens if I overwrite the start of a LUKS partition or
damage the LUKS header or key-slots?
There are two critical components for decryption: The salt values in the
key-slot descriptors of the header and the key-slots. For LUKS2 they
are a bit better protected. but for LUKS1, these are right in the first
sector. If the salt values are overwritten or changed, nothing (in the
cryptographically strong sense) can be done to access the data, unless
there is a backup of the LUKS header. If a key-slot is damaged, the
data can still be read with a different key-slot, if there is a
remaining undamaged and used key-slot. Note that in order to make a
key-slot completely unrecoverable, changing about 4-6 bits in random
locations of its 128kiB size is quite enough.
* 6.9 What happens if I (quick) format a LUKS partition?
I have not tried the different ways to do this, but very likely you will
have written a new boot-sector, which in turn overwrites the LUKS
header, including the salts, making your data permanently irretrievable,
unless you have a LUKS header backup. For LUKS2 this may still be
recoverable without that header backup, for LUKS1 it is not. You may
also damage the key-slots in part or in full. See also last item.
* 6.10 How do I recover the master key from a mapped LUKS1 container?
Note: LUKS2 uses the kernel keyring to store keys and hence this
procedure does not work unless you have explicitly disabled the use of
the keyring with "--disable-keyring" on opening.
This is typically only needed if you managed to damage your LUKS1
header, but the container is still mapped, i.e. "luksOpen"ed. It also
helps if you have a mapped container that you forgot or do not know a
passphrase for (e.g. on a long running server.)
WARNING: Things go wrong, do a full backup before trying this!
WARNING: This exposes the master key of the LUKS1 container. Note that
both ways to recreate a LUKS header with the old master key described
below will write the master key to disk. Unless you are sure you have
securely erased it afterwards, e.g. by writing it to an encrypted
partition, RAM disk or by erasing the filesystem you wrote it to by a
complete overwrite, you should change the master key afterwards.
Changing the master key requires a full data backup, luksFormat and then
restore of the backup. Alternatively the tool cryptsetup-reencrypt from
the cryptsetup package can be used to change the master key (see its
man-page), but a full backup is still highly recommended.
First, there is a script by Milan that automates the whole process,
except generating a new LUKS1 header with the old master key (it prints
the command for that though):
https://gitlab.com/cryptsetup/cryptsetup/blob/master/misc/luks-header-from-active
You can also do this manually. Here is how:
- Get the master key from the device mapper. This is done by the
following command. Substitute c5 for whatever you mapped to:
# dmsetup table --target crypt --showkey /dev/mapper/c5
Result:
0 200704 crypt aes-cbc-essiv:sha256
a1704d9715f73a1bb4db581dcacadaf405e700d591e93e2eaade13ba653d0d09
0 7:0 4096
The result is actually one line, wrapped here for clarity. The long
hex string is the master key.
- Convert the master key to a binary file representation. You can do
this manually, e.g. with hexedit. You can also use the tool "xxd"
from vim like this:
echo "a1704d9....53d0d09" | xxd -r -p > <master-key-file>
- Do a luksFormat to create a new LUKS1 header.
NOTE: If your header is intact and you just forgot the passphrase,
you can just set a new passphrase, see next sub-item.
Unmap the device before you do that (luksClose). Then do
cryptsetup luksFormat --master-key-file=<master-key-file> <luks device>
Note that if the container was created with other than the default
settings of the cryptsetup version you are using, you need to give
additional parameters specifying the deviations. If in doubt, try the
script by Milan. It does recover the other parameters as well.
Side note: This is the way the decrypt_derived script gets at the master
key. It just omits the conversion and hashes the master key string.
- If the header is intact and you just forgot the passphrase, just
set a new passphrase like this:
cryptsetup luksAddKey --master-key-file=<master-key-file> <luks device>
You may want to disable the old one afterwards.
* 6.11 What does the on-disk structure of dm-crypt look like?
There is none. dm-crypt takes a block device and gives encrypted access
to each of its blocks with a key derived from the passphrase given. If
you use a cipher different than the default, you have to specify that as
a parameter to cryptsetup too. If you want to change the password, you
basically have to create a second encrypted device with the new
passphrase and copy your data over. On the plus side, if you
accidentally overwrite any part of a dm-crypt device, the damage will be
limited to the area you overwrote.
* 6.12 What does the on-disk structure of LUKS1 look like?
Note: For LUKS2, refer to the LUKS2 document referenced in Item 1.2
A LUKS1 partition consists of a header, followed by 8 key-slot
descriptors, followed by 8 key slots, followed by the encrypted data
area.
Header and key-slot descriptors fill the first 592 bytes. The key-slot
size depends on the creation parameters, namely on the number of
anti-forensic stripes, key material offset and master key size.
With the default parameters, each key-slot is a bit less than 128kiB in
size. Due to sector alignment of the key-slot start, that means the key
block 0 is at offset 0x1000-0x20400, key block 1 at offset
0x21000-0x40400, and key block 7 at offset 0xc1000-0xe0400. The space
to the next full sector address is padded with zeros. Never used
key-slots are filled with what the disk originally contained there, a
key-slot removed with "luksRemoveKey" or "luksKillSlot" gets filled with
0xff. Due to 2MiB default alignment, start of the data area for
cryptsetup 1.3 and later is at 2MiB, i.e. at 0x200000. For older
versions, it is at 0x101000, i.e. at 1'052'672 bytes, i.e. at 1MiB +
4096 bytes from the start of the partition. Incidentally,
"luksHeaderBackup" for a LUKS container created with default parameters
dumps exactly the first 2MiB (or 1'052'672 bytes for headers created
with cryptsetup versions < 1.3) to file and "luksHeaderRestore" restores
them.
For non-default parameters, you have to figure out placement yourself.
"luksDump" helps. See also next item. For the most common non-default
settings, namely aes-xts-plain with 512 bit key, the offsets are: 1st
keyslot 0x1000-0x3f800, 2nd keyslot 0x40000-0x7e000, 3rd keyslot
0x7e000-0xbd800, ..., and start of bulk data at 0x200000.
The exact specification of the format is here:
https://gitlab.com/cryptsetup/cryptsetup/wikis/Specification
For your convenience, here is the LUKS1 header with hex offsets.
NOTE:
The spec counts key-slots from 1 to 8, but the cryptsetup tool counts
from 0 to 7. The numbers here refer to the cryptsetup numbers.
Refers to LUKS1 On-Disk Format Specification Version 1.2.3
LUKS1 header:
offset length name data type description
-----------------------------------------------------------------------
0x0000 0x06 magic byte[] 'L','U','K','S', 0xba, 0xbe
0 6
0x0006 0x02 version uint16_t LUKS version
6 3
0x0008 0x20 cipher-name char[] cipher name spec.
8 32
0x0028 0x20 cipher-mode char[] cipher mode spec.
40 32
0x0048 0x20 hash-spec char[] hash spec.
72 32
0x0068 0x04 payload-offset uint32_t bulk data offset in sectors
104 4 (512 bytes per sector)
0x006c 0x04 key-bytes uint32_t number of bytes in key
108 4
0x0070 0x14 mk-digest byte[] master key checksum
112 20 calculated with PBKDF2
0x0084 0x20 mk-digest-salt byte[] salt for PBKDF2 when
132 32 calculating mk-digest
0x00a4 0x04 mk-digest-iter uint32_t iteration count for PBKDF2
164 4 when calculating mk-digest
0x00a8 0x28 uuid char[] partition UUID
168 40
0x00d0 0x30 key-slot-0 key slot key slot 0
208 48
0x0100 0x30 key-slot-1 key slot key slot 1
256 48
0x0130 0x30 key-slot-2 key slot key slot 2
304 48
0x0160 0x30 key-slot-3 key slot key slot 3
352 48
0x0190 0x30 key-slot-4 key slot key slot 4
400 48
0x01c0 0x30 key-slot-5 key slot key slot 5
448 48
0x01f0 0x30 key-slot-6 key slot key slot 6
496 48
0x0220 0x30 key-slot-7 key slot key slot 7
544 48
Key slot:
offset length name data type description
-------------------------------------------------------------------------
0x0000 0x04 active uint32_t key slot enabled/disabled
0 4
0x0004 0x04 iterations uint32_t PBKDF2 iteration count
4 4
0x0008 0x20 salt byte[] PBKDF2 salt
8 32
0x0028 0x04 key-material-offset uint32_t key start sector
40 4 (512 bytes/sector)
0x002c 0x04 stripes uint32_t number of anti-forensic
44 4 stripes
* 6.13 What is the smallest possible LUKS1 container?
Note: From cryptsetup 1.3 onwards, alignment is set to 1MB. With modern
Linux partitioning tools that also align to 1MB, this will result in
alignment to 2k sectors and typical Flash/SSD sectors, which is highly
desirable for a number of reasons. Changing the alignment is not
recommended.
That said, with default parameters, the data area starts at exactly 2MB
offset (at 0x101000 for cryptsetup versions before 1.3). The smallest
data area you can have is one sector of 512 bytes. Data areas of 0
bytes can be created, but fail on mapping.
While you cannot put a filesystem into something this small, it may
still be used to contain, for example, key. Note that with current
formatting tools, a partition for a container this size will be 3MiB
anyways. If you put the LUKS container into a file (via losetup and a
loopback device), the file needs to be 2097664 bytes in size, i.e. 2MiB
+ 512B.
The two ways to influence the start of the data area are key-size and
alignment.
For alignment, you can go down to 1 on the parameter. This will still
leave you with a data-area starting at 0x101000, i.e. 1MiB+4096B
(default parameters) as alignment will be rounded up to the next
multiple of 8 (i.e. 4096 bytes) If in doubt, do a dry-run on a larger
file and dump the LUKS header to get actual information.
For key-size, you can use 128 bit (e.g. AES-128 with CBC), 256 bit
(e.g. AES-256 with CBC) or 512 bit (e.g. AES-256 with XTS mode). You
can do 64 bit (e.g. blowfish-64 with CBC), but anything below 128 bit
has to be considered insecure today.
Example 1 - AES 128 bit with CBC:
cryptsetup luksFormat -s 128 --align-payload=8 <device>
This results in a data offset of 0x81000, i.e. 516KiB or 528384
bytes. Add one 512 byte sector and the smallest LUKS container size
with these parameters is 516KiB + 512B or 528896 bytes.
Example 2 - Blowfish 64 bit with CBC (WARNING: insecure):
cryptsetup luksFormat -c blowfish -s 64 --align-payload=8 /dev/loop0
This results in a data offset of 0x41000, i.e. 260kiB or 266240
bytes, with a minimal LUKS1 container size of 260kiB + 512B or 266752
bytes.
* 6.14 I think this is overly complicated. Is there an alternative?
Not really. Encryption comes at a price. You can use plain dm-crypt to
simplify things a bit. It does not allow multiple passphrases, but on
the plus side, it has zero on disk description and if you overwrite some
part of a plain dm-crypt partition, exactly the overwritten parts are
lost (rounded up to full sectors).
* 6.15 Can I clone a LUKS container?
You can, but it breaks security, because the cloned container has the
same header and hence the same master key. Even if you change the
passphrase(s), the master key stays the same. That means whoever has
access to one of the clones can decrypt them all, completely bypassing
the passphrases.
While you can use cryptsetup-reencrypt to change the master key,
this is probably more effort than to create separate LUKS containers
in the first place.
The right way to do this is to first luksFormat the target container,
then to clone the contents of the source container, with both containers
mapped, i.e. decrypted. You can clone the decrypted contents of a LUKS
container in binary mode, although you may run into secondary issues
with GUIDs in filesystems, partition tables, RAID-components and the
like. These are just the normal problems binary cloning causes.
Note that if you need to ship (e.g.) cloned LUKS containers with a
default passphrase, that is fine as long as each container was
individually created (and hence has its own master key). In this case,
changing the default passphrase will make it secure again.
7. Interoperability with other Disk Encryption Tools
* 7.1 What is this section about?
Cryptsetup for plain dm-crypt can be used to access a number of on-disk
formats created by tools like loop-aes patched into losetup. This
sometimes works and sometimes does not. This section collects insights
into what works, what does not and where more information is required.
Additional information may be found in the mailing-list archives,
mentioned at the start of this FAQ document. If you have a solution
working that is not yet documented here and think a wider audience may
be interested, please email the FAQ maintainer.
* 7.2 loop-aes: General observations.
One problem is that there are different versions of losetup around.
loop-aes is a patch for losetup. Possible problems and deviations
from cryptsetup option syntax include:
- Offsets specified in bytes (cryptsetup: 512 byte sectors)
- The need to specify an IV offset
- Encryption mode needs specifying (e.g. "-c twofish-cbc-plain")
- Key size needs specifying (e.g. "-s 128" for 128 bit keys)
- Passphrase hash algorithm needs specifying
Also note that because plain dm-crypt and loop-aes format does not have
metadata, and while the loopAES extension for cryptsetup tries
autodetection (see command loopaesOpen), it may not always work. If you
still have the old set-up, using a verbosity option (-v) on mapping with
the old tool or having a look into the system logs after setup could
give you the information you need. Below, there are also some things
that worked for somebody.
* 7.3 loop-aes patched into losetup on Debian 5.x, kernel 2.6.32
In this case, the main problem seems to be that this variant of
losetup takes the offset (-o option) in bytes, while cryptsetup takes
it in sectors of 512 bytes each.
Example: The losetup command
losetup -e twofish -o 2560 /dev/loop0 /dev/sdb1
mount /dev/loop0 mount-point
translates to
cryptsetup create -c twofish -o 5 --skip 5 e1 /dev/sdb1
mount /dev/mapper/e1 mount-point
* 7.4 loop-aes with 160 bit key
This seems to be sometimes used with twofish and blowfish and represents
a 160 bit ripemed160 hash output padded to 196 bit key length. It seems
the corresponding options for cryptsetup are
--cipher twofish-cbc-null -s 192 -h ripemd160:20
* 7.5 loop-aes v1 format OpenSUSE
Apparently this is done by older OpenSUSE distros and stopped working
from OpenSUSE 12.1 to 12.2. One user had success with the following:
cryptsetup create <target> <device> -c aes -s 128 -h sha256
* 7.6 Kernel encrypted loop device (cryptoloop)
There are a number of different losetup implementations for using
encrypted loop devices so getting this to work may need a bit of
experimentation.
NOTE: Do NOT use this for new containers! Some of the existing
implementations are insecure and future support is uncertain.
Example for a compatible mapping:
losetup -e twofish -N /dev/loop0 /image.img
translates to
cryptsetup create image_plain /image.img -c twofish-cbc-plain -H plain
with the mapping being done to /dev/mapper/image_plain instead of
to /dev/loop0.
More details:
Cipher, mode and password hash (or no hash):
-e cipher [-N] => -c cipher-cbc-plain -H plain [-s 256]
-e cipher => -c cipher-cbc-plain -H ripemd160 [-s 256]
Key size and offsets (losetup: bytes, cryptsetuop: sectors of 512 bytes):
-k 128 => -s 128
-o 2560 => -o 5 -p 5 # 2560/512 = 5
There is no replacement for --pass-fd, it has to be emulated using
keyfiles, see the cryptsetup man-page.
8. Issues with Specific Versions of cryptsetup
* 8.1 When using the create command for plain dm-crypt with
cryptsetup 1.1.x, the mapping is incompatible and my data is not
accessible anymore!
With cryptsetup 1.1.x, the distro maintainer can define different
default encryption modes. You can check the compiled-in defaults using
"cryptsetup --help". Moreover, the plain device default changed because
the old IV mode was vulnerable to a watermarking attack.
If you are using a plain device and you need a compatible mode, just
specify cipher, key size and hash algorithm explicitly. For
compatibility with cryptsetup 1.0.x defaults, simple use the following:
cryptsetup create -c aes-cbc-plain -s 256 -h ripemd160 <name> <dev>
LUKS stores cipher and mode in the metadata on disk, avoiding this
problem.
* 8.2 cryptsetup on SLED 10 has problems...
SLED 10 is missing an essential kernel patch for dm-crypt, which is
broken in its kernel as a result. There may be a very old version of
cryptsetup (1.0.x) provided by SLED, which should also not be used
anymore as well. My advice would be to drop SLED 10.
* 8.3 Gcrypt 1.6.x and later break Whirlpool
It is the other way round: In gcrypt 1.5.x, Whirlpool is broken and it
was fixed in 1.6.0 and later. If you selected whirlpool as hash on
creation of a LUKS container, it does not work anymore with the fixed
library. This shows one serious risk of using rarely used settings.
Note that at the time this FAQ item was written, 1.5.4 was the latest
1.5.x version and it has the flaw, i.e. works with the old Whirlpool
version. Possibly later 1.5.x versions will work as well. If not,
please let me know.
The only two ways to access older LUKS containers created with Whirlpool
are to either decrypt with an old gcrypt version that has the flaw or to
use a compatibility feature introduced in cryptsetup 1.6.4 and gcrypt
1.6.1 or later. Version 1.6.0 cannot be used.
Steps:
- Make at least a header backup or better, refresh your full backup.
(You have a full backup, right? See Item 6.1 and following.)
- Make sure you have cryptsetup 1.6.4 or later and check the gcrypt
version:
cryptsetup luksDump <your luks device> --debug | grep backend
If gcrypt is at version 1.5.x or before:
- Reencrypt the LUKS header with a different hash. (Requires entering
all keyslot passphrases. If you do not have all, remove the ones you
do not have before.):
cryptsetup-reencrypt --keep-key --hash sha256 <your luks device>
If gcrypt is at version 1.6.1 or later:
- Patch the hash name in the LUKS header from "whirlpool" to
"whirlpool_gcryptbug". This activates the broken implementation.
The detailed header layout is in Item 6.12 of this FAQ and in the
LUKS on-disk format specification. One way to change the hash is
with the following command:
echo -n -e 'whirlpool_gcryptbug\0' | dd of=<luks device> bs=1 seek=72 conv=notrunc
- You can now open the device again. It is highly advisable to change
the hash now with cryptsetup-reencrypt as described above. While you
can reencrypt to use the fixed whirlpool, that may not be a good idea
as almost nobody seems to use it and hence the long time until the
bug was discovered.
9. The Initrd question
* 9.1 My initrd is broken with cryptsetup
That is not nice! However the initrd is supplied by your distribution,
not by the cryptsetup project and hence you should complain to them. We
cannot really do anything about it.
* 9.2 CVE-2016-4484 says cryptsetup is broken!
Not really. It says the initrd in some Debian versions have a behavior
that under some very special and unusual conditions may be considered
a vulnerability.
What happens is that you can trick the initrd to go to a rescue-shell if
you enter the LUKS password wrongly in a specific way. But falling back
to a rescue shell on initrd errors is a sensible default behavior in the
first place. It gives you about as much access as booting a rescue
system from CD or USB-Stick or as removing the disk would give you. So
this only applies when an attacker has physical access, but cannot boot
anything else or remove the disk. These will be rare circumstances
indeed, and if you rely on the default distribution initrd to keep you
safe under these circumstances, then you have bigger problems than this
somewhat expected behavior.
The CVE was exagerrated and should not be assigned to upstream
cryptsetup in the first place (it is a distro specific initrd issue).
It was driven more by a try to make a splash for self-aggrandizement,
than by any actual security concerns. Ignore it.
* 9.3 How do I do my own initrd with cryptsetup?
Note: The instructions here apply to an initrd in initramfs format, not
to an initrd in initrd format. The latter is a filesystem image, not a
cpio-archive, and seems to not be widely used anymore.
It depends on the distribution. Below, I give a very simple example and
step-by-step instructions for Debian. With a bit of work, it should be
possible to adapt this to other distributions. Note that the
description is pretty general, so if you want to do other things with an
initrd it provides a useful starting point for that too.
01) Unpacking an existing initrd to use as template
A Linux initrd is in gzip'ed cpio format. To unpack it, use something
like this:
md tmp; cd tmp; cat ../initrd | gunzip | cpio -id
After this, you have the full initrd content in tmp/
02) Inspecting the init-script
The init-script is the only thing the kernel cares about. All activity
starts there. Its traditional location is /sbin/init on disk, but /init
in an initrd. In an initrd unpacked as above it is tmp/init.
While init can be a binary despite usually being called "init script",
in Debian the main init on the root partition is a binary, but the init
in the initrd (and only that one is called by the kernel) is a script
and starts like this:
#!/bin/sh
....
The "sh" used here is in tmp/bin/sh as just unpacked, and in Debian it
currently is a busybox.
03) Creating your own initrd
The two examples below should give you most of what is needed. This is
tested with LUKS1 and should work with LUKS2 as well. If not, please
let me know.
Here is a really minimal example. It does nothing but set up some
things and then drop to an interactive shell. It is perfect to try out
things that you want to go into the init-script.
#!/bin/sh
export PATH=/sbin:/bin
[ -d /sys ] || mkdir /sys
[ -d /proc ] || mkdir /proc
[ -d /tmp ] || mkdir /tmp
mount -t sysfs -o nodev,noexec,nosuid sysfs /sys
mount -t proc -o nodev,noexec,nosuid proc /proc
echo "initrd is running, starting BusyBox..."
exec /bin/sh --login
Here is an example that opens the first LUKS-partition it finds with the
hard-coded password "test2" and then mounts it as root-filesystem. This
is intended to be used on an USB-stick that after boot goes into a safe,
as it contains the LUKS-passphrase in plain text and is not secure to be
left in the system. The script contains debug-output that should make it
easier to see what is going on. Note that the final hand-over to the init
on the encrypted root-partition is done by "exec switch_root /mnt/root
/sbin/init", after mounting the decrypted LUKS container with "mount
/dev/mapper/c1 /mnt/root". The second argument of switch_root is relative
to the first argument, i.e. the init started with this command is really
/mnt/sbin/init before switch_root runs.
#!/bin/sh
export PATH=/sbin:/bin
[ -d /sys ] || mkdir /sys
[ -d /proc ] || mkdir /proc
[ -d /tmp ] || mkdir /tmp
mount -t sysfs -o nodev,noexec,nosuid sysfs /sys
mount -t proc -o nodev,noexec,nosuid proc /proc
echo "detecting LUKS containers in sda1-10, sdb1-10"; sleep 1
for i in a b
do
for j in 1 2 3 4 5 6 7 8 9 10
do
sleep 0.5
d="/dev/sd"$i""$j
echo -n $d
cryptsetup isLuks $d >/dev/null 2>&1
r=$?
echo -n " result: "$r""
# 0 = is LUKS, 1 = is not LUKS, 4 = other error
if expr $r = 0 > /dev/null
then
echo " is LUKS, attempting unlock"
echo -n "test2" | cryptsetup luksOpen --key-file=- $d c1
r=$?
echo " result of unlock attempt: "$r""
sleep 2
if expr $r = 0 > /dev/null
then
echo "*** LUKS partition unlocked, switching root ***
echo " (waiting 30 seconds before doing that)"
mount /dev/mapper/c1 /mnt/root
sleep 30
exec switch_root /mnt/root /sbin/init
fi
else
echo " is not LUKS"
fi
done
done
echo "FAIL finding root on LUKS, loading BusyBox..."; sleep 5
exec /bin/sh --login
04) What if I want a binary in the initrd, but libraries are missing?
That is a bit tricky. One option is to compile statically, but that
does not work for everything. Debian puts some libraries into lib/ and
lib64/ which are usually enough. If you need more, you can add the
libraries you need there. That may or may not need a configuration
change for the dynamic linker "ld" as well. Refer to standard Linux
documentation on how to add a library to a Linux system. A running
initrd is just a running Linux system after all, it is not special in
any way.
05) How do I repack the initrd?
Simply repack the changed directory. While in tmp/, do
the following:
```
find . | cpio --create --format='newc' | gzip > ../new_initrd
```
Rename "new_initrd" to however you want it called (the name of
the initrd is a kernel-parameter) and move to /boot. That is it.
10. LUKS2 Questions
* 10.1 Is the cryptography of LUKS2 different?
Mostly not. The header has changed in its structure, but the
crytpgraphy is the same. The one exception is that PBKDF2 has been
replaced by Argon2 to give better resilience against attacks attacks by
graphics cards and other hardware with lots of computing power but
limited local memory per computing element.
* 10.2 What new features does LUKS2 have?
There are quite a few. I recommend reading the man-page and the on-disk
format specification, see Item 1.2.
To list just some:
- A lot of the metadata is JSON, allowing for easier extension
- Max 32 key-slots per default
- Better protection for bad passphrases now available with Argon2
- Authenticated encryption
- The LUKS2 header is less vulnerable to corruption and has a 2nd copy
* 10.3 Why does LUKS2 need so much memory?
LUKS2 uses Argon2 instead of PBKDF2. That causes the increase in memory.
See next item.
* 10.4 Why use Argon2 in LUKS 2 instead of PBKDF2?
LUKS tries to be secure with not-so-good passwords. Bad passwords need to
be protected in some way against an attacker that just tries all possible
combinations. (For good passwords, you can just wait for the attacker to
die of old age...) The situation with LUKS is not quite the same as with a
password stored in a database, but there are similarities.
LUKS does not store passwords on disk. Instead, the passwords are used to
decrypt the master-key with it and that one is stored on disk in encrypted
form. If you have a good password, with, say, more than 80 bits of
entropy, you could just put the password through a single crypto-hash (to
turn it into something that can be used as a key) and that would be secure.
This is what plain dm-crypt does.
If the password has lower entropy, you want to make this process cost some
effort, so that each try takes time and resources and slows the attacker
down. LUKS1 uses PBKDF2 for that, adding an iteration count and a salt.
The iteration count is per default set to that it takes 1 second per try on
the CPU of the device where the respective passphrase was set. The salt is
there to prevent precomputation.
The problem with that is that if you use a graphics card, you can massively
speed up these computations as PBKDF2 needs very little memeory to compute
it. A graphics card is (grossly simplified) a mass of small CPUs with some
small very fast local memory per CPU and a large slow memory (the 4/6/8 GB
a current card may have). If you can keep a computation in the small,
CPU-local memory, you can gain a speed factor of 1000 or more when trying
passwords with PBKDF2.
Argon2 was created to address this problem. It adds a "large memory
property" where computing the result with less memory than the memory
parameter requires is massively (exponentially) slowed down. That means,
if you set, for example, 4GB of memory, computing Argon2 on a graphics card
with around 100kB of memory per "CPU" makes no sense at all because it is
far too slow. An attacker has hence to use real CPUs and furthermore is
limited by main memory bandwith.
Hence the large amount of memory used is a security feature and should not
be turned off or reduced. If you really (!) understand what you are doing
and can assure good passwords, you can either go back to PBKDF2 or set a
low amount of memory used for Argon2 when creating the header.
* 10.5 LUKS2 is insecure! It uses less memory than the Argon2 RFC say!
Well, not really. The RFC recommends 6GiB of memory for use with disk
encryption. That is a bit insane and something clearly went wrong in the
standardization process here. First, that makes Argon2 unusable on any 32
bit Linux and that is clearly a bad thing. Second, there are many small
Linux devices around that do not have 6GiB of RAM in the first place. For
example, the current Raspberry Pi has 1GB, 2GB or 4GB of RAM, and with the
RFC recommendations, none of these could compute Argon2 hashes.
Hence LUKS2 uses a more real-world approach. Iteration is set to a
minimum of 4 because there are some theoretical attacks that work up to an
iteration count of 3. The thread parameter is set to 4. To achieve 2
second/slot unlock time, LUKS2 adjusts the memory parameter down if
needed. In the other direction, it will respect available memory and not
exceed it. On a current PC, the memory parameter will be somewhere around
1GB, which should quite generous. The minimum I was able to set in an
experiment with "-i 1" was 400kB of memory and that is too low to be
secure. A Raspberry Pi would probably end up somewhere around 50MB (have
not tried it) and that should still be plenty.
That said, if you have a good, high-entropy passphrase, LUKS2 is secure
with any memory parameter.
* 10.6 How does re-encryption store data while it is running?
All metadata necessary to perform a recovery of said segment (in case of
crash) is stored in the LUKS2 metadata area. No matter if the LUKS2
reencryption was run in online or offline mode.
* 10.7 What do I do if re-encryption crashes?
In case of a reencryption application crash, try to close the original
device via following command first:
cryptsetup close <my_crypt_device>.
Cryptsetup assesses if it's safe to teardown the reencryption device stack
or not. It will also cut off I/O (via dm-error mapping) to current
hotzone segment (to make later recovery possible). If it can't be torn
down, i.e. due to a mounted fs, you must unmount the filesystem first.
Never try to tear down reencryption dm devices manually using e.g.
dmsetup tool, at least not unless cryptsetup says it's safe to do so. It
could damage the data beyond repair.
* 10.8 Do I need to enter two passphrases to recover a crashed
re-encryption?
Cryptsetup (command line utility) expects the passphrases to be identical
for the keyslot containing old volume key and for the keyslot containing
new one. So the recovery happens during normal the "cryptsetup open"
operation or the equivalent during boot.
Re-encryption recovery can be also performed in offline mode by
the "cryptsetup repair" command.
* 10.9 What is an unbound keyslot and what is it used for?
Quite simply, an 'unbound key' is an independent 'key' stored in a luks2
keyslot that cannot be used to unlock a LUKS2 data device. More specifically,
an 'unbound key' or 'unbound luks2 keyslot' contains a secret that is not
currently associated with any data/crypt segment (encrypted area) in the
LUKS2 'Segments' section (displayed by luksDump).
This is a bit of a more general idea. It basically allows to use a keyslot
as a container for a key to be used in other things than decrypting a
data segment.
As of April 2020, the following uses are defined:
1) LUKS2 re-encryption. The new volume key is stored in an unbound keyslot
which becomes a regular LUKS2 keyslot later when re-encryption is
finished.
2) Somewhat similar is the use with a wrapped key scheme (e.g. with the
paes cipher). In this case, the VK (Volume Key) stored in a keyslot
is an encrypted binary binary blob. The KEK (Key Encryption Key) for
that blob may be refreshed (Note that this KEK is not managed by
cryptsetup!) and the binary blob gets changed. The KEK refresh process
uses an 'unbound keyslot'. First the future effective VK is placed
in the unbound keyslot and later it gets turned into the new real VK
(and bound to the respective crypt segment).
* 10.10 What about the size of the LUKS2 header?
While the LUKS1 header has a fixed size that is determined by the cipher
spec (see Item 6.12), LUKS2 is more variable. The default size is 16MB,
but it can be adjusted on creation by using the --luks2-metadata-size
and --luks2-keyslots-size options. Refer to the man-page for details.
While adjusting the size in an existing LUKS2 container is possible,
it is somewhat complicated and risky. My advice is to do a backup,
recreate the container with changed parameters and restore that backup.
* 10.11 Does LUKS2 store metadata anywhere except in the header?
It does not. But note that if you use the experimental integrity support,
there will be an integrity header as well at the start of the data area
and things get a bit more complicated. All metadata will still be at the
start of the device, nothing gets stored somewhere in the middle or at
the end.
11. References and Further Reading
* Purpose of this Section
The purpose of this section is to collect references to all materials
that do not fit the FAQ but are relevant in some fashion. This can be
core topics like the LUKS spec or disk encryption, but it can also be
more tangential, like secure storage management or cryptography used in
LUKS. It should still have relevance to cryptsetup and its
applications.
If you want to see something added here, send email to the maintainer
(or the cryptsetup mailing list) giving an URL, a description (1-3 lines
preferred) and a section to put it in. You can also propose new
sections.
At this time I would like to limit the references to things that are
available on the web.
* Specifications
- LUKS on-disk format spec: See Item 1.2
* Other Documentation
- Arch Linux on LUKS, LVM and full-disk encryption:
https://wiki.archlinux.org/index.php/Dm-crypt/Encrypting_an_entire_system
* Code Examples
- Some code examples are in the source package under docs/examples
- LUKS AF Splitter in Ruby by John Lane: https://rubygems.org/gems/afsplitter
* Brute-forcing passphrases
- http://news.electricalchemy.net/2009/10/password-cracking-in-cloud-part-5.html
- http://it.slashdot.org/story/12/12/05/0623215/new-25-gpu-monster-devours-strong-passwords-in-minutes
* Tools
* SSD and Flash Disk Related
* Disk Encryption
* Attacks Against Disk Encryption
* Risk Management as Relevant for Disk Encryption
* Cryptography
* Secure Storage
A. Contributors
In no particular order:
- Arno Wagner
- Milan Broz
___