8 6. Backup and Data Recovery
9 7. Issues with Specific Versions of cryptsetup
18 This is the FAQ (Frequently Asked Questions) for cryptsetup. It
19 covers Linux disk encryption with plain dm-crypt (one passphrase,
20 no management, no descriptor on disk) and LUKS (multiple user keys
21 with one master key, anti-forensics, descriptor block at start of
22 device, ...). The latest version should usually be available at
23 http://code.google.com/p/cryptsetup/wiki/FrequentlyAskedQuestions
25 ATTENTION: If you are going to read just one thing, make it the
26 section on Backup and Data Recovery. By far the most questions on
27 the cryptsetup mailing list are from people that just managed to
28 somehow format or overwrite the start of their LUKS partitions. In
29 most cases, there is nothing that can be done to help these poor
30 souls recover their data. Make sure you understand the problem and
31 limitations imposed by the LUKS security model BEFORE you face such
37 Current FAQ maintainer is Arno Wagner <arno@wagner.name>. Other
38 contributors are listed at the end. If you want to contribute, send
39 your article, including a descriptive headline, to the maintainer,
40 or the dm-crypt mailing list with something like "FAQ ..." in the
41 subject. Please note that by contributing to this FAQ, you accept
42 the license described below.
44 This work is under the "Attribution-Share Alike 3.0 Unported"
45 license, which means distribution is unlimited, you may create
46 derived works, but attributions to original authors and this
47 license statement must be retained and the derived work must be
48 under the same license. See
49 http://creativecommons.org/licenses/by-sa/3.0/ for more details of
52 Side note: I did text license research some time ago and I think
53 this license is best suited for the purpose at hand and creates the
57 * Where is the project website?
59 There is the project website at http://code.google.com/p/cryptsetup/
60 Please do not post questions there, nobody will read them. Use
61 the mailing-list instead.
64 * Is there a mailing-list?
66 Instructions on how to subscribe to the mailing-list are at on the
67 project website. People are generally helpful and friendly on the
70 The question of how to unsubscribe from the list does crop up
71 sometimes. For this you need your list management URL, which is
72 sent to you initially and once at the start of each month. Go to
73 the URL mentioned in the email and select "unsubscribe". This page
74 also allows you to request a password reminder.
76 Alternatively, you can send an Email to dm-crypt-request@saout.de
77 with just the word "help" in the subject or message body. Make sure
78 to send it from your list address.
80 The mailing list archive is here:
81 http://dir.gmane.org/gmane.linux.kernel.device-mapper.dm-crypt
87 * What is the difference between "plain" and LUKS format?
89 Plain format is just that: It has no metadata on disk, reads all
90 paramters from the commandline (or the defaults), derives a
91 master-key from the passphrase and then uses that to de-/encrypt
92 the sectors of the device, with a direct 1:1 mapping between
93 encrypted and decrypted sectors.
95 Primary advantage is high resilience to damage, as one damaged
96 encrypted sector results in exactly one damaged decrypted sector.
97 Also, it is not readily apparent that there even is encrypted data
98 on the device, as an overwrite with crypto-grade randomness (e.g.
99 from /dev/urandom) looks exactly the same on disk.
101 Side-note: That has limited value against the authorities. In
102 civilized countries, they cannot force you to give up a crypto-key
103 anyways. In the US, the UK and dictatorships around the world,
104 they can force you to give up the keys (using imprisonment or worse
105 to pressure you), and in the worst case, they only need a
106 nebulous "suspicion" about the presence of encrypted data. My
107 advice is to either be ready to give up the keys or to not have
108 encrypted data when traveling to those countries, especially when
109 crossing the borders.
111 Disadvantages are that you do not have all the nice features that
112 the LUKS metadata offers, like multiple passphrases that can be
113 changed, the cipher being stored in the metadata, anti-forensic
114 properties like key-slot diffusion and salts, etc..
116 LUKS format uses a metadata header and 8 key-slot areas that are
117 being placed ath the begining of the disk, see below under "What
118 does the LUKS on-disk format looks like?". The passphrases are used
119 to decryt a single master key that is stored in the anti-forensic
122 Advantages are a higher usability, automatic configuration of
123 non-default crypto parameters, defenses against low-entropy
124 passphrases like salting and iterated PBKDF2 passphrase hashing,
125 the ability to change passhrases, and others.
127 Disadvantages are that it is readily obvious there is encrypted
128 data on disk (but see side note above) and that damage to the
129 header or key-slots usually results in permanent data-loss. See
130 below under "6. Backup and Data Recovery" on how to reduce that
131 risk. Also the sector numbers get shifted by the length of the
132 header and key-slots and there is a loss of that size in capacity
133 (1MB+4096B for defaults and 2MB for the most commonly used
134 non-default XTS mode).
137 * Can I encrypt an already existing, non-empty partition to use LUKS?
139 There is no converter, and it is not really needed. The way to do
140 this is to make a backup of the device in question, securely wipe
141 the device (as LUKS device initialization does not clear away old
142 data), do a luksFormat, optionally overwrite the encrypted device,
143 create a new filesystem and restore your backup on the now
144 encrypted device. Also refer to sections "Security Aspects" and
145 "Backup and Data Recovery".
147 For backup, plain GNU tar works well and backs up anything likely
148 to be in a filesystem.
151 * How do I use LUKS with a loop-device?
153 Just the same as with any block device. If you want, for example,
154 to use a 100MiB file as LUKS container, do something like this:
156 head -c 100M /dev/zero > luksfile # create empty file
157 losetup /dev/loop0 luksfile # map luksfile to /dev/loop0
158 cryptsetup luksFormat /dev/loop0 # create LUKS on loop device
160 Afterwards just use /dev/loop0 as a you would use a LUKS partition.
161 To unmap the file when done, use "losetup -d /dev/loop0".
164 * When I add a new key-slot to LUKS, it asks for a passphrase but
165 then complains about there not being a key-slot with that
168 That is as intended. You are asked a passphrase of an existing
169 key-slot first, before you can enter the passphrase for the new
170 key-slot. Otherwise you could break the encryption by just adding a
171 new key-slot. This way, you have to know the passphrase of one of
172 the already configured key-slots in order to be able to configure a
176 * How do I read a dm-crypt key from file?
178 Note that the file will still be hashed first, just like keyboard
179 input. Use the --key-file option, like this:
181 cryptsetup create --key-file keyfile e1 /dev/loop0
184 * How do I read a LUKS slot key from file?
186 What you really do here is to read a passphrase from file, just as
187 you would with manual entry of a passphrase for a key-slot. You can
188 add a new passphrase to a free key-slot, set the passphrase of an
189 specific key-slot or put an already configured passphrase into a
190 file. In the last case make sure no trailing newline (0x0a) is
191 contained in the key file, or the passphrase will not work because
192 the whole file is used as input.
194 To add a new passphrase to a free key slot from file, use something
197 cryptsetup luksAddKey /dev/loop0 keyfile
199 To add a new passphrase to a specific key-slot, use something like
202 cryptsetup luksAddKey --key-slot 7 /dev/loop0 keyfile
204 To supply a key from file to any LUKS command, use the --key-file
205 option, e.g. like this:
207 cryptsetup luksOpen --key-file keyfile /dev/loop0 e1
210 * How do I read the LUKS master key from file?
212 The question you should ask yourself first is why you would want to
213 do this. The only legitimate reason I can think of is if you want
214 to have two LUKS devices with the same master key. Even then, I
215 think it would be preferable to just use key-slots with the same
216 passphrase, or to use plain dm-crypt instead. If you really have a
217 good reason, please tell me. If I am convinced, I will add how to
221 * What are the security requirements for a key read from file?
223 A file-stored key or passphrase has the same security requirements
224 as one entered interactively, however you can use random bytes and
225 thereby use bytes you cannot type on the keyboard. You can use any
226 file you like as key file, for example a plain text file with a
227 human readable passphrase. To generate a file with random bytes,
228 use something like this:
230 head -c 256 /dev/random > keyfile
233 * If I map a journaled file system using dm-crypt/LUKS, does it still
234 provide its usual transactional guarantees?
236 As far as I know it does (but I may be wrong), but please note that
237 these "guarantees" are far weaker than they appear to be. For
238 example, you may not get a hard flush to disk surface even on a
239 call to fsync. In addition, the HDD itself may do independent
240 write reordering. Some other things can go wrong as well. The
241 filesystem developers are aware of these problems and typically
242 can make it work anyways. That said, dm-crypt/LUKS should not make
245 Personally, I have several instances of ext3 on dm-crypt and have
246 not noticed any specific problems.
248 Update: I did run into frequent small freezes (1-2 sec) when putting
249 a vmware image on ext3 over dm-crypt. This does indicate that the
250 transactional guarantees are in place, but at a cost. When I went
251 back to ext2, the problem went away. This also seems to have gotten
252 better with kernel 2.6.36 and the reworking of filesystem flush
253 locking. Kernel 2.6.38 is expected to have more improvements here.
256 * Can I use LUKS or cryptsetup with a more secure (external) medium
257 for key storage, e.g. TPM or a smartcard?
259 Yes, see the answers on using a file-supplied key. You do have to
260 write the glue-logic yourself though. Basically you can have
261 cryptsetup read the key from STDIN and write it there with your
262 own tool that in turn gets the key from the more secure key
266 * Can I resize a dm-crypt or LUKS partition?
268 Yes, you can, as neither dm-crypt nor LUKS stores partition size.
269 Whether you should is a different question. Personally I recommend
270 backup, recreation of the encrypted partition with new size,
271 recreation of the filesystem and restore. This gets around the
272 tricky business of resizing the filesystem. Resizing a dm-crypt or
273 LUKS container does not resize the filesystem in it. The backup is
274 really non-optional here, as a lot can go wrong, resulting in
275 partial or complete data loss. Using something like gparted to
276 resize an encrypted partition is slow, but typicaly works. This
277 will not change the size of the filesystem hidden under the
280 You also need to be aware of size-based limitations. The one
281 currently relevant is that aes-xts-plain should not be used for
282 encrypted container sizes larger than 2TiB. Use aes-xts-plain64
289 * My dm-crypt/LUKS mapping does not work! What general steps are
290 there to investigate the problem?
292 If you get a specific error message, investigate what it claims
293 first. If not, you may want to check the following things.
295 - Check that "/dev", including "/dev/mapper/control" is there. If it
296 is missing, you may have a problem with the "/dev" tree itself or
297 you may have broken udev rules.
299 - Check that you have the device mapper and the crypt target in your
300 kernel. The output of "dmsetup targets" should list a "crypt"
301 target. If it is not there or the command fails, add device mapper
302 and crypt-target to the kernel.
304 - Check that the hash-functions and ciphers you want to use are in
305 the kernel. The output of "cat /proc/crypto" needs to list them.
308 * My dm-crypt mapping suddenly stopped when upgrading cryptsetup.
310 The default cipher, hash or mode may have changed (the mode changed
311 from 1.0.x to 1.1.x). See under "Issues With Specific Versions of
315 * When I call cryptsetup from cron/CGI, I get errors about unknown
318 If you get errors about unknown parameters or the like that are not
319 present when cryptsetup is called from the shell, make sure you
320 have no older version of cryptsetup on your system that then gets
321 called by cron/CGI. For example some distributions install
322 cryptsetup into /usr/sbin, while a manual install could go to
323 /usr/local/sbin. As a debugging aid, call "cryptsetup --version"
324 from cron/CGI or the non-shell mechanism to be sure the right
328 * Unlocking a LUKS device takes very long. Why?
330 The iteration time for a key-slot (see Section 5 for an explanation
331 what iteration does) is calculated when setting a passphrase. By
332 default it is 1 second on the machine where the passphrase is set.
333 If you set a passphrase on a fast machine and then unlock it on a
334 slow machine, the unlocking time can be much longer. Also take into
335 account that up to 8 key-slots have to be tried in order to find the
338 If this is problem, you can add another key-slot using the slow
339 machine with the same passphrase and then remove the old key-slot.
340 The new key-slot will have an iteration count adjusted to 1 second
341 on the slow machine. Use luksKeyAdd and then luksKillSlot or
344 However, this operation will not change volume key iteration count
345 (MK iterations in output of "cryptsetup luksDump"). In order to
346 change that, you will have to backup the data in the LUKS
347 container, luksFormat on the slow machine and restore the data.
348 Note that in the original LUKS specification this value was fixed
349 to 10, but it is now derived from the PBKDF2 benchmark as well and
350 set to iterations in 0.125 sec or 1000, whichever is larger.
353 * "blkid" sees a LUKS UUID and an ext2/swap UUID on the same device.
356 Some old versions of cryptsetup have a bug where the header does
357 not get completely wiped during LUKS format and an older ext2/swap
358 signature remains on the device. This confuses blkid.
360 Fix: Wipe the unused header areas by doing a backup and restore of
361 the header with cryptsetup 1.1.x:
363 cryptsetup luksHeaderBackup --header-backup-file <file> <device>
364 cryptsetup luksHeaderRestore --header-backup-file <file> <device>
367 * cryptsetup segfaults on Gentoo amd64 hardened ...
369 There seems to be some inteference between the hardening and and
370 the way cryptsetup benchmarks PBKDF2. The solution to this is
371 currently not quite clear for an encrypted root filesystem. For
372 other uses, you can apparently specify USE="dynamic" as compile
373 flag, see http://bugs.gentoo.org/show_bug.cgi?id=283470
379 * Can a bad RAM module cause problems?
381 LUKS and dm-crypt can give the RAM quite a workout, especially when
382 combined with software RAID. In particular the combination RAID5 +
383 LUKS + XFS seems to uncover RAM problems that never caused obvious
384 problems before. Symptoms vary, but often the problem manifest
385 itself when copying large amounts of data, typically several times
386 larger than your main memory.
388 Side note: One thing you should always do on large data
389 copy/movements is to run a verify, for example with the "-d"
390 option of "tar" or by doing a set of MD5 checksums on the source
393 find . -type f -exec md5sum \{\} \; > checksum-file
395 and then a "md5sum -c checksum-file" on the other side. If you get
396 mismatches here, RAM is the primary suspect. A lesser suspect is
397 an overclocked CPU. I have found countless hardware problems in
398 verify runs after copying or making backups. Bit errors are much
399 more common than most people think.
401 Some RAM issues are even worse and corrupt structures in one of the
402 layers. This typically results in lockups, CPU state dumps in the
403 system logs, kernel panic or other things. It is quite possible to
404 have the problem with an encrypted device, but not with an
405 otherwise the same unencrypted device. The reason for that is that
406 encryption has an error amplification property: You flip one bit
407 in an encrypted data block, and the decrypted version has half of
408 its bits flipped. This is an important security property for modern
409 ciphers. With the usual modes in cryptsetup (CBC, ESSIV, XTS), you
410 get up to a completely changed 512 byte block per bit error. A
411 corrupt block causes a lot more havoc than the occasionally
412 flipped single bit and can result various obscure errors.
414 Note however that a verify run on copying between encrypted or
415 unencrypted devices can also show you corruption when the copying
416 itself did not report any problems. If you find defect RAM, assume
417 all backups and copied data to be suspect, unless you did a verify.
422 First you should know that overclocking often makes memory
423 problems worse. So if you overclock (which I strongly recommend
424 against in a system holding data that has some worth), run the
425 tests with the overclocking active.
427 There are two good options. One is Memtest86+ and the other is
428 "memtester" by Charles Cazabon. Memtest86+ requires a reboot and
429 then takes over the machine, while memtester runs from a
430 root-shell. Both use different testing methods and I have found
431 problems fast with each one that the other needed long to find. I
432 recommend running the following procedure until the first error is
435 - Run Memtest86+ for one cycle
437 - Run memterster for one cycle (shut down as many other applications
440 - Run Memtest86+ for 24h or more
442 - Run memtester for 24h or more
444 If all that does not produce error messages, your RAM may be sound,
445 but I have had one weak bit that Memtest86+ needed around 60 hours
446 to find. If you can reproduce the original problem reliably, a good
447 additional test may be to remove half of the RAM (if you have more
448 than one module) and try whether the problem is still there and if
449 so, try with the other half. If you just have one module, get a
450 different one and try with that. If you do overclocking, reduce
451 the settings to the most conservative ones available and try with
458 * Should I initialize (overwrite) a new LUKS/dm-crypt partition?
460 If you just create a filesystem on it, most of the old data will
461 still be there. If the old data is sensitive, you should overwrite
462 it before encrypting. In any case, not initializing will leave the
463 old data there until the specific sector gets written. That may
464 enable an attacker to determine how much and where on the
465 partition data was written. If you think this is a risk, you can
466 prevent this by overwriting the encrypted device (here assumed to
467 be named "e1") with zeros like this:
469 dd_rescue -w /dev/zero /dev/mapper/e1
471 or alternatively with one of the following more standard commands:
473 cat /dev/zero > /dev/mapper/e1
474 dd if=/dev/zero of=/dev/mapper/e1
477 * How do I securely erase a LUKS (or other) partition?
479 For LUKS, if you are in a desperate hurry, overwrite the LUKS
480 header and key-slot area. This means overwriting the first
481 (keyslots x stripes x keysize) + offset bytes. For the default
482 parameters, this is the 1'052'672 bytes, i.e. 1MiB + 4096 of the
483 LUKS partition. For 512 bit key length (e.g. for aes-xts-plain with
484 512 bit key) this is 2MiB. (The diferent offset stems from
485 differences in the sector alignment of the key-slots.) If in doubt,
486 just be generous and overwrite the first 10MB or so, it will likely
487 still be fast enough. A single overwrite with zeros should be
488 enough. If you anticipate being in a desperate hurry, prepare the
489 command beforehand. Example with /dev/sde1 as the LUKS partition
490 and default parameters:
492 head -c 1052672 /dev/zero > /dev/sde1; sync
494 A LUKS header backup or full backup will still grant access to
495 most or all data, so make sure that an attacker does not have
496 access to backups or destroy them as well.
498 If you have time, overwrite the whole LUKS partition with a single
499 pass of zeros. This is enough for current HDDs. For SSDs or FLASH
500 (USB sticks) you may want to overwrite the whole drive several
501 times to be sure data is not retained by wear leveling. This is
502 possibly still insecure as SSD technology is not fully understood
503 in this regard. Still, due to the anti-forensic properties of the
504 LUKS key-slots, a single overwrite of an SSD or FLASH drive could
505 be enough. If in doubt, use physical destruction in addition. Here
506 is a link to some current reseach results on erasing SSDs and FLASH
508 http://www.usenix.org/events/fast11/tech/full_papers/Wei.pdf
510 Keep in mind to also erase all backups.
512 Example for a zero-overwrite erase of partition sde1 done with
515 dd_rescue -w /dev/zero /dev/sde1
518 * How do I securely erase a backup of a LUKS partition or header?
520 That depends on the medium it is stored on. For HDD and SSD, use
521 overwrite with zeros. For an SSD or FLASH drive (USB stick), you
522 may want to overwrite the complete SSD several times and use
523 physical destruction in addition, see last item. For re-writable
524 CD/DVD, a single overwrite should also be enough, due to the
525 anti-forensic properties of the LUKS keyslots. For write-once
526 media, use physical destruction. For low security requirements,
527 just cut the CD/DVD into several parts. For high security needs,
528 shred or burn the medium. If your backup is on magnetic tape, I
529 advise physical destruction by shredding or burning, after
530 overwriting . The problem with magnetic tape is that it has a
531 higher dynamic range than HDDs and older data may well be
532 recoverable after overwrites. Also write-head alignment issues can
533 lead to data not actually being deleted at all during overwrites.
536 * What about backup? Does it compromise security?
538 That depends. See next section.
541 * Why is all my data permanently gone if I overwrite the LUKS header?
543 Overwriting the LUKS header in part or in full is the most common
544 reason why access to LUKS containers is lost permanently.
545 Overwriting can be done in a number of fashions, like creating a
546 new filesystem on the raw LUKS partition, making the raw partition
547 part of a raid array and just writing to the raw partition.
549 The LUKS header contains a 256 bit "salt" value and without that no
550 decryption is possible. While the salt is not secret, it is
551 key-grade material and cannot be reconstructed. This is a
552 cryptographically strong "cannot". From observations on the
553 cryptsetup mailing-list, people typically go though the usual
554 stages of grief (Denial, Anger, Bargaining, Depression, Acceptance)
555 when this happens to them. Observed times vary between 1 day and 2
556 weeks to complete the cycle. Seeking help on the mailing-list is
557 fine. Even if we usually cannot help with getting back your data,
558 most people found the feedback comforting.
560 If your header does not contain an intact salt, best go directly
561 to the last stage ("Acceptance") and think about what to do now.
562 There is one exception that I know of: If your LUKS container is
563 still open, then it may be possible to extract the master key from
564 the running system. Ask on the mailing-list on how to do that and
565 make sure nobody switches off the machine.
570 A salt is a random key-grade value added to the passphrase before
571 it is processed. It is not kept secret. The reason for using salts
572 is as follows: If an attacker wants to crack the password for a
573 single LUKS container, then every possible passphrase has to be
574 tried. Typically an attacker will not try every binary value, but
575 will try words and sentences from a dictionary.
577 If an attacker wants to attack several LUKS containers with the
578 same dictionary, then a different approach makes sense: Compute the
579 resulting slot-key for each dictionary element and store it on
580 disk. Then the test for each entry is just the slow unlocking with
581 the slot key (say 0.00001 sec) instead of calculating the slot-key
582 first (1 sec). For a single attack, this does not help. But if you
583 have more than one container to attack, this helps tremendously,
584 also because you can prepare your table before you even have the
585 container to attack! The calculation is also very simple to
586 parallelize. You could, for example, use the night-time unused CPU
587 power of your desktop PCs for this.
589 This is where the salt comes in. If the salt is combined with the
590 passphrase (in the simplest form, just appended to it), you
591 suddenly need a separate table for each salt value. With a
592 reasonably-sized salt value (256 bit, e.g.) this is quite
596 * Is LUKS secure with a low-entropy (bad) passphrase?
598 This needs a bit of theory. The quality of your passphrase is
599 directly related to its entropy (information theoretic, not
600 thermodynamic). The entropy says how many bits of "uncertainty" or
601 "randomness" are in you passphrase. In other words, that is how
602 difficult guessing the passphrase is.
604 Example: A random English sentence has about 1 bit of entropy per
605 character. A random lowercase (or uppercase) character has about
608 Now, if n is the number of bits of entropy in your passphrase and t
609 is the time it takes to process a passphrase in order to open the
610 LUKS container, then an attacker has to spend at maximum
612 attack_time_max = 2^n * t
614 time for a successful attack and on average half that. There is no
615 way getting around that relationship. However, there is one thing
616 that does help, namely increasing t, the time it takes to use a
617 passphrase, see next FAQ item.
619 Still, if you want good security, a high-entropy passphrase is the
620 only option. Use at least 64 bits for secret stuff. That is 64
621 characters of English text (but only if randomly chosen) or a
622 combination of 12 truly random letters and digits.
624 For passphrase generation, do not use lines from very well-known
625 texts (religious texts, Harry potter, etc.) as they are to easy to
626 guess. For example, the total Harry Potter has about 1'500'000
627 words (my estimation). Trying every 64 character sequence starting
628 and ending at a word boundary would take only something like 20
629 days on a single CPU and is entirely feasible. To put that into
630 perspective, using a number of Amazon EC2 High-CPU Extra Large
631 instances (each gives about 8 real cores), this tests costs
632 currently about $48, but can be made to run arbitrarily fast.
634 On the other hand, choosing 1.5 lines from, say, the Wheel of Time
635 is in itself not more secure, but the book selection adds quite a
636 bit of entropy. (Now that I have mentioned it here, don't use tWoT
637 either!) If you add 2 or 3 typos or switch some words around, then
638 this is good passphrase material.
641 * What is "iteration count" and why is decreasing it a bad idea?
643 Iteration count is the number of PBKDF2 iterations a passphrase is
644 put through before it is used to unlock a key-slot. Iterations are
645 done with the explicit purpose to increase the time that it takes
646 to unlock a key-slot. This provides some protection against use of
647 low-entropy passphrases.
649 The idea is that an attacker has to try all possible passphrases.
650 Even if the attacker knows the passphrase is low-entropy (see last
651 item), it is possible to make each individual try take longer. The
652 way to do this is to repeatedly hash the passphrase for a certain
653 time. The attacker then has to spend the same time (given the same
654 computing power) as the user per try. With LUKS, the default is 1
655 second of PBKDF2 hashing.
657 Example 1: Lets assume we have a really bad passphrase (e.g. a
658 girlfriends name) with 10 bits of entropy. With the same CPU, an
659 attacker would need to spend around 500 seconds on average to
660 break that passphrase. Without iteration, it would be more like
661 0.0001 seconds on a modern CPU.
663 Example 2: The user did a bit better and has 32 chars of English
664 text. That would give use about 32 bits of entropy. With 1 second
665 iteration, that means an attacker on the same CPU needs around 136
666 years. That is pretty impressive for such a weak passphrase.
667 Without the iterations, it would be more like 50 days on a modern
668 CPU, and possibly far less.
670 In addition, the attacker can both parallelize and use special
671 hardware like GPUs to speed up the attack. The attack can also
672 happen quite some time after the luksFormat operation and CPUs can
673 have become faster and cheaper. For that reason you want a bit
674 of extra security. Anyways, in Example 1 your are screwed. In
675 example 2, not necessarily. Even if the attack is faster, it still
676 has a certain cost associated with it, say 10000 EUR/USD with
677 iteration and 1 EUR/USD without iteration. The first can be
678 prohibitively expensive, while the second is something you try
679 even without solid proof that the decryption will yield something
682 The numbers above are mostly made up, but show the idea. Of course
683 the best thing is to have a high-entropy passphrase.
685 Would a 100 sec iteration time be even better? Yes and no.
686 Cryptographically it would be a lot better, namely 100 times better.
687 However, usability is a very important factor for security
688 technology and one that gets overlooked surprisingly often. For
689 LUKS, if you have to wait 2 minutes to unlock the LUKS container,
690 most people will not bother and use less secure storage instead. It
691 is better to have less protection against low-entropy passphrases
692 and people actually use LUKS, than having them do without
693 encryption altogether.
695 Now, what about decreasing the iteration time? This is generally a
696 very bad idea, unless you know and can enforce that the users only
697 use high-entropy passphrases. If you decrease the iteration time
698 without ensuring that, then you put your users at increased risk,
699 and considering how rarely LUKS containers are unlocked in a
700 typical work-flow, you do so without a good reason. Don't do it.
701 The iteration time is already low enough that users with entropy
702 low passphrases are vulnerable. Lowering it even further increases
703 this danger significantly.
706 * Is LUKS with default parameters less secure on a slow CPU?
708 Unfortunately, yes. However the only aspect affected is the
709 protection for low-entropy passphrase or master-key. All other
710 security aspects are independent of CPU speed.
712 The master key is less critical, as you really have to work at it
713 to give it low entropy. One possibility is to supply the master key
714 yourself. If that key is low-entropy, then you get what you
715 deserve. The other known possibility is to use /dev/urandom for
716 key generation in an entropy-startved situation (e.g. automatic
717 installation on an embedded device without network and other entropy
720 For the passphrase, don't use a low-entropy passphrase. If your
721 passphrase is good, then a slow CPU will not matter. If you insist
722 on a low-entropy passphrase on a slow CPU, use something like
723 "--iter-time=10" or higher and wait a long time on each LUKS unlock
724 and pray that the attacker does not find out in which way exactly
725 your passphrase is low entropy. This also applies to low-entropy
726 passphrases on fast CPUs. Technology can do only so much to
727 compensate for problems in front of the keyboard.
730 * Why was the default aes-cbc-plain replaced with aes-cbc-essiv?
732 The problem is that cbc-plain has a fingerprint vulnerability, where
733 a specially crafted file placed into the crypto-container can be
734 recognized from the outside. The issue here is that for cbc-plain
735 the initialization vector (IV) is the sector number. The IV gets
736 XORed to the first data chunk of the sector to be encrypted. If you
737 make sure that the first data block to be stored in a sector
738 contains the sector number as well, the first data block to be
739 encrypted is all zeros and always encrypted to the same ciphertext.
740 This also works if the first data chunk just has a constant XOR
741 with the sector number. By having several shifted patterns you can
742 take care of the case of a non-power-of-two start sector number of
745 This mechanism allows you to create a pattern of sectors that have
746 the same first ciphertext block and signal one bit per sector to the
747 outside, allowing you to e.g. mark media files that way for
748 recognition without decryption. For large files this is a
749 practical attack. For small ones, you do not have enough blocks to
750 signal and take care of different file starting offsets.
752 In order to prevent this attack, the default was changed to
753 cbc-essiv. ESSIV uses a keyed hash of the sector number, with the
754 encryption key as key. This makes the IV unpredictable without
755 knowing the encryption key and the watermarking attack fails.
758 * Are there any problems with "plain" IV? What is "plain64"?
760 First, "plain" and "plain64" are both not secure to use with CBC,
761 see previous FAQ item.
763 However there are modes, like XTS, that are secure with "plain" IV.
764 The next limit is that "plain" is 64 bit, with the upper 32 bit set
765 to zero. This means that on volumes larger than 2TiB, the IV
766 repeats, creating a vulnerability that potentially leaks some
767 data. To avoid this, use "plain64", which uses the full sector
768 number up to 64 bit. Note that "plain64" requires a kernel >=
769 2.6.33. Also note that "plain64" is backwards compatible for
770 volume sizes <= 2TiB, but not for those > 2TiB. Finally, "plain64"
771 does not cause any performance penalty compared to "plain".
774 * What about XTS mode?
776 XTS mode is potentially even more secure than cbc-essiv (but only if
777 cbc-essiv is insecure in your scenario). It is a NIST standard and
778 used, e.g. in Truecrypt. At the moment, if you want to use it, you
779 have to specify it manually as "aes-xts-plain", i.e.
781 cryptsetup -c aes-xts-plain luksFormat <device>
783 For volumes >2TiB and kernels >= 2.6.33 use "plain64" (see FAQ
784 item on "plain" and "plain64"):
786 cryptsetup -c aes-xts-plain64 luksFormat <device>
788 There is a potential security issue with XTS mode and large blocks.
789 LUKS and dm-crypt always use 512B blocks and the issue does not
793 6. Backup and Data Recovery
796 * Does a backup compromise security?
798 Depends on how you do it. First, a backup is non-optional with
799 encrypted data just the same way it is with non-encrypted data.
800 Disks do break and they do not care whether they make plain or
801 encrypted data inaccessible. As a gideline, a well-treated HDD (!)
802 breaks with about 5% probability per year. This means everybody
803 will be hit sooner or later.
805 However there are risks introduced by backups. For example if you
806 change/disable a key-slot in LUKS, a binary backup of the partition
807 will still have the old key-slot. To deal with this, you have to
808 be able to change the key-slot on the backup as well, or use a
809 different set-up. One option is to have a different passphrase on
810 the backup and to make the backup with both containers open.
811 Another one is to make a backup of the original, opened container
812 to a single file, e.g. with tar, and to encrypt that file with
813 public-key-cryptography, e.g. with GnuPG. You can then keep the
814 secret key in a safe place, because it is only used to decrypt a
815 backup. The key the backup is encrypted with can be stored without
816 special security measures, as long as an attacker cannot replace
819 If you use dm-crypt, backup is simpler: As there is no key
820 management, the main risk is that you cannot wipe the backup when
821 wiping the original. However wiping the original for dm-crypt
822 should consist of forgetting the passphrase and that you can do
823 without actual access to the backup.
825 In both cases, there is an additional (usually small) risk: An
826 attacker can see how many sectors and which ones have been changed
827 since the backup. This is not possible with the public-key method
830 My personal advice is to use one USB disk (low value date) or
831 three disks (high value data) in rotating order for backups, and
832 either use different passphrases or keep them easily accessible
833 in case you need to disable a key-slot. If you do network-backup
834 or tape-backup, I strongly recommend to go the public-key path,
835 especially as you typically cannot reliably delete data in these
836 scenarios. (Well, you can burn the tape if it is under your
840 * What happens if I overwrite the start of a LUKS partition or damage
841 the LUKS header or key-slots?
843 There are two critical components for decryption: The salt values
844 in the header itself and the key-slots. If the salt values are
845 overwritten or changed, nothing (in the cryptographically strong
846 sense) can be done to access the data, unless there is a backup
847 of the LUKS header. If a key-slot is damaged, the data can still
848 be read with a different key-slot, if there is a remaining
849 undamaged and used key-slot. Note that in order to make a key-slot
850 unrecoverable in a cryptographically strong sense, changing about
851 4-6 bits in random locations of its 128kiB size is quite enough.
854 * What happens if I (quick) format a LUKS partition?
856 I have not tried the different ways to do this, but very likely you
857 will have written a new boot-sector, which in turn overwrites the
858 LUKS header, including the salts, making your data permanently
859 irretrivable, unless you have a LUKS header backup. You may also
860 damage the key-slots in part or in full. See also last item.
863 * What does the on-disk structure of dm-crypt look like?
865 There is none. dm-crypt takes a block device and gives encrypted
866 access to each of its blocks with a key derived from the passphrase
867 given. If you use a cipher different than the default, you have to
868 specify that as a parameter to cryptsetup too. If you want to
869 change the password, you basically have to create a second
870 encrypted device with the new passphrase and copy your data over.
871 On the plus side, if you accidentally overwrite any part of a
872 dm-crypt device, the damage will be limited to the are you
876 * What does the on-disk structure of LUKS look like?
878 A LUKS partition consists of a header, followed by 8 key-slot
879 descriptors, followed by 8 key slots, followed by the encrypted
882 Header and key-slot descriptors fill the first 592 bytes. The
883 key-slot size depends on the creation parameters, namely on the
884 number of anti-forensic stripes, key material offset and master
887 With the default parameters, each key-slot is a bit less than
888 128kiB in size. Due to sector alignment of the key-slot start,
889 that means the key block 0 is at offset 0x1000-0x20400, key
890 block 1 at offset 0x21000-0x40400, and key block 7 at offset
891 0xc1000-0xe0400. The space to the next full sector address is
892 padded with zeros. Never used key-slots are filled with what the
893 disk originally contained there, a key-slot removed with
894 "luksRemoveKey" or "luksKillSlot" gets filled with 0xff. Start of
895 bulk data is at 0x101000, i.e. at 1'052'672 bytes, i.e. at 1MiB
896 + 4096 bytes from the start of the partition. This is also the
897 value given by command "luksDump" with "Payload offset: 2056",
898 just multiply by the sector size (512 bytes). Incidentally,
899 "luksHeaderBackup" for a LUKS container created with default
900 parameters dumps exactly the first 1'052'672 bytes to file and
901 "luksHeaderRestore" restores them.
903 For non-default parameters, you have to figure out placement
904 yourself. "luksDump" helps. For the most common non-default
905 settings, namely aes-xts-plain with 512 bit key, the offsets are:
906 1st keyslot 0x1000-0x3f800, 2nd keyslot 0x40000-0x7e000, 3rd
907 keyslot 0x7e000-0xbd800, ..., and start of bulk data at 0x200000.
909 The exact specification of the format is here:
910 http://code.google.com/p/cryptsetup/wiki/Specification
913 * How do I backup a LUKS header?
915 While you could just copy the appropriate number of bytes from the
916 start of the LUKS partition, the best way is to use command option
917 "luksHeaderBackup" of cryptsetup. This protects also against
918 errors when non-standard parameters have been used in LUKS
919 partition creation. Example:
922 cryptsetup luksHeaderBackup --header-backup-file h /dev/mapper/c1
924 To restore, use the inverse command, i.e.
926 cryptsetup luksHeaderRestore --header-backup-file h /dev/mapper/c1
929 * How do I backup a LUKS partition?
931 You do a sector-image of the whole partition. This will contain
932 the LUKS header, the keys-slots and the data ares. It can be done
933 under Linux e.g. with dd_rescue (for a direct image copy) and with
934 "cat" or "dd". Example:
936 cat /dev/sda10 > sda10.img
937 dd_rescue /dev/sda10 sda10.img
939 You can also use any other backup software that is capable of making
940 a sector image of a partition. Note that compression is
941 ineffective for encrypted data, hence it does not make sense to
945 * Do I need a backup of the full partition? Would the header and
946 key-slots not be enough?
948 Backup protects you against two things: Disk loss or corruption
949 and user error. By far the most questions on the dm-crypt mailing
950 list about how to recover a damaged LUKS partition are related
951 to user error. For example, if you create a new filesystem on a
952 LUKS partition, chances are good that all data is lost
955 For this case, a header+key-slot backup would often be enough. But
956 keep in mind that a well-treated (!) HDD has roughly a failure
957 risk of 5% per year. It is highly advisable to have a complete
958 backup to protect against this case.
961 * Are there security risks from a backup of the LUKS header or a
962 whole LUKS partition?
964 Yes. One risk is that if you remove access rights for specific
965 key-slots by deleting their contents, the data can still be
966 accessed with invalidated passphrase and the backup. The other
967 risk is that if you erase a LUKS partition, a backup could still
968 grant access, especially if you only erased the LUKS header and
969 not the whole partition.
972 * I think this is overly complicated. Is there an alternative?
974 Yes, you can use plain dm-crypt. It does not allow multiple
975 passphrases, but on the plus side, it has zero on disk description
976 and if you overwrite some part of a plain dm-crypt partition,
977 exactly the overwritten parts are lost (rounded up to sector
981 7. Issues with Specific Versions of cryptsetup
984 * When using the create command for plain dm-crypt with cryptsetup
985 1.1.x, the mapping is incompatible and my data is not accessible
988 With cryptsetup 1.1.x, the distro maintainer can define different
989 default encryption modes for LUKS and plain devices. You can check
990 these compiled-in defaults using "cryptsetup --help". Moreover, the
991 plain device default changed because the old IV mode was
992 vulnerable to a watermarking attack.
994 If you are using a plain device and you need a compatible mode, just
995 specify cipher, key size and hash algorithm explicitly. For
996 compatibility with cryptsetup 1.0.x defaults, simple use the
999 cryptsetup create -c aes-cbc-plain -s 256 -h ripemd160 <name> <dev>
1001 LUKS stores cipher and mode in the metadata on disk, avoiding this
1005 * cryptsetup on SLED 10 has problems...
1007 SLED 10 is missing an essential kernel patch for dm-crypt, which
1008 is broken in its kernel as a result. There may be a very old
1009 version of cryptsetup (1.0.x) provided by SLED, which should also
1010 not be used anymore as well. My advice would be to drop SLED 10.
1012 A. Contributors In no particular order: