Re: [openpgp] Dealing with clock skew

["Derek Atkins" <[e-mail]>]

Hi Neal,

There is plenty of prior art on dealing with time skew.  Kerberos, for
example, has been dealing with it since the 1980s!  In short, you have a
sliding window of acceptance.  Kerberos sets this as +/- 5 minutes.  So if
the timestamp is within that 10-minute range it is considered current
(valid).

If you are seeing 20-minute clock skew on a VM, then you should really set
up ntp.

As for your second issue, my questions to you are:
1) are the messages are guaranteed to be in order (or could they be
reordered).
2) can the validator maintain per-sender state?

If YES to both questions, then the easiest thing to do is maintain the
last timestamp sent and reject if the next message has a timestamp prior
to the last one seen.  Note that this state must be per-sender, so if you
have Alice talking to Bob and Charlie, Alice needs to maintain separate
last-seen numbers for both Bob and Charlie, because their clocks can be
skewed differently.

Hope this helps,

-derek

On Fri, November 15, 2019 6:17 am, Neal H. Walfield wrote:
> Hi,
>
> Our OpenPGP library is being used in an application that does a
> real-time exchange.  Specifically, two MUAs sometimes communicate via
> an imap folder.  During testing, the application developers observed
> that the message exchange would sometimes fail, because a signature
> was considered invalid.
>
> Eventually, we figured out that the problem was due to clock skew: the
> sender would sign a message at, say, 9:00 and the receiver would
> correctly reject it because its clock said that the current time was
> 8:59, and 9:00, the time stored in the message's Signature Creation
> Time subpacket, is clearly in the future.
>
> It turns out that a second of clock skew is actually pretty good.  The
> testing was done on a VM and the application developer reported that
> up to 20 minutes of clock skew are not unusual.
>
> Adding a tolerance seems like a reasonable way to deal with this
> problem.  But, there are situations where a tolerance would cause
> other problems.  Consider, for instance, an application that generates
> messages that contain its current state, and we want to know its state
> at some point in time.  Specifically, consider messages, Ma and Mb,
> and a time T.  If Ta < T < Tb (< current time), then we should use Ma.
> But, if T is significantly close to Tb, we may not reject Mb and use
> it instead!
>
> An obvious example of this type of message is self signatures.  It is
> tempting to just not use tolerances with self-signatures, but if an
> application generates a key, and immediately sends it to another
> computer, the second computer may reject the key as being invalid,
> because it doesn't consider it to have any valid self-signatures!  So,
> some tolerance is also needed for self signatures.  Also, the example
> also applies to normal ("user data") messages, so specially handling
> self-signatures is not the right approach, anyway.
>
> Perhaps the best thing to do is to only use a tolerance when asking if
> a signature is valid "now".
>
>
> How do other implementations deal with this?  Thoughts?
>
> Thanks!
>
> :) Neal
>
> _______________________________________________
> openpgp mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/openpgp
>


-- 
       Derek Atkins                 617-623-3745
       [e-mail]             www.ihtfp.com
       Computer and Internet Security Consultant