[dm-crypt] security: improve defaults

Milan Broz gmazyland at gmail.com
Sat Jan 3 10:08:22 CET 2015


Adding just fey my notes...

On 01/03/2015 07:00 AM, Arno Wagner wrote:
> On Sat, Jan 03, 2015 at 00:18:21 CET, Christian Stadelmann wrote:

>> cipher: aes-cbc-essiv (default in plain mode)
>> There are known attacs against aes-cbc-essiv which lead to using aes-xts
>> as default cipher in LUKS mode. Is there any reason why it should not be
>> used in plain mode?
> 
> Simple: Backwards compatibility. As plain mode does not
> have a header, this would break old uses. Anybody that wants
> it can already use XTS.

BTW XTS has own flaws, the main problem is that granularity of a change
(small plaintext change propagates only to one cipher block in ciphertext).
But unfortunately we do not have better mode (for the default).
Very nice comments are in
http://csrc.nist.gov/groups/ST/toolkit/BCM/documents/comments/XTS/collected_XTS_comments.pdf

And definitely the CBC should be phased out but, for plain
mode the compatibility is the main problem - you will break many people
systems when switching default.

>> key size: 256 (default)
>> For using aes256 (which is the default cipher in LUKS mode) the key size
>> should be 512 bit since XTS splits the supplied key.
> 
> The default is AES-128. If you have some insights why/how that
> is insecure, please cite them.  

Fully agree.

>> hash: sha1 (default)
>> SHA-1 is considered weak for some years, SHA-2 is widely available. Is
>> there any reason against using SHA-2? Since hashing is only done once
>> sha512 could be default.
> 
> Not for this use, see FAQ Item 5.20.

This will change once other LUKS header get new major version (probably together
with PBKDF change, see below). There is no problem with SHA1 here but
it will save many these useless "SHA1" discussions at least ...

>  
>> iter-time: 1000 (default)
>> could be increased.
> 
> Or it could not. It is a well balanced default, that leads to an 
> unlock time of up to 10 seconds and an unlock fail time of 10
> seconds. Increasing it by default would be a problem in many cases.

Yes, 1 second is a nice balanced default, if you are unlocking the last keyslot
and all slots are used, you have 8 seconds unlocking time!
Actually others often request to decrease this default ;-)

>> random number pool: /dev/urandom (default)
>> this should definitely be `--use-random` as default, you should never
>> use /dev/urandom for long-term crypto keys. It may result in using
>> low-entropy keys which obviously must not happen. It might take some
>> time to gather enough entropy, but that is ok since performance is not
>> relevant for an operation done once. Additionaly I think it would be
>> best to disable the option `--use-urandom` completely.
> 
> No. Really not. There is extensive discussion of this question on 
> the mailing-list. Read that and then come again.

The same story as SHA1... I already planned to switch upstream to /dev/random
as a default just to stop wasting time on this issue.

Please read the discussions, it is not as simple as it seems.
Let's start with http://www.2uo.de/myths-about-urandom/

We should care more about proper entropy to initially seed RNG.
If you think you have a problem, install some external device
seeding your RNG. 

>> key derivation function: PBKDF2
>> PBKDF2 is easy to implement in FPGAs or ASICs which reduces its
>> strength. It is safe enough for today but scrypt is a good alternative. 
> 
> There is a password-hashing competition going one at this time:
> 
>   https://password-hashing.net/
> 
> It would be pretty stupid not waiting for the results.
> Also, scrypt cancelled its standardization attempt 
> (the IETF RFC-draft expired), so it is not really in a 
> final  state. Most of the entries in the competition are 
> also much better than scrypt and it is reasonable to 
> expect that the winners will complete standardization
> and then can form a solid base for version 2.0 of
> the LUKS header.

I can promise that LUKS will be one of the first users of PHC output.
I already spent a lot of time playing with these hasning functions
so no need to worry:-)

There is the excellent yescrypt as a direct evolution to scrypt
as one of the 2nd round candidate.

But really, nobody will use it now for production when there is
no proper analysis of these algorithms and candidates are not yet even chosen.

And one note to PBKDF2/LUKS:

Yes, PBKDF2 itself can run very fast on GPU/ASIC/FPGA but did you try LUKS?
The AF filter spreads out keyslot to quite large sector area, and you
must have this area in memory to perform attack. Nice side effect which
complicates LUKS attack on these platforms.

> 
>> To summarize: Strong crypto is available. It should be default.
> 
> That reasoning is flawed. The primary requirement for crypto
> software is usability. Not used crypto can be extremely
> secure without that helping one bit, as not used software 
> protects noting. Sure, security is a strong second requirement,
> but you have failed to demonstrate any insecurty that is
> not trumped by usability. 

Fully agree. Cryptsetup allows to configure crypto according to your
needs. If your analysis says you need something better than default,
just use it! But forcing other users to use "stronger" crypto
could cause exactly opposite - incompatible containers which cannot
be opened on older system, slow unlocking, ...
Many users will stop to use it once this happens. So it is about
finding usability tradeoff here.

Milan



More information about the dm-crypt mailing list