For Developers‎ > ‎Design Documents‎ > ‎

Time Sources

Reliably figuring out what time it is on a machine operating in an untrustworthy environment is pretty difficult. Currently we use the following sources:
  1. The machine's realtime clock, if set
  2. htpdate(1) pointed at
Unfortunately, at first boot, the realtime clock may not be set well, and there exist platforms where the realtime clock does not persist across power cycles because it is not battery-backed. We need to get the right time on these machines, mainly to avoid attacks on TLS connections (by setting the date far enough backwards, an attacker might render a TLS certificate that expired valid again). Right now, htpdate(1) refuses to set the date to before the time it was built. There could also be attacks that are opened up by pushing the system clock far enough forward (e.g., causing various system daemons to see files on disk with an apparent modification time far in the future, or causing cryptohomes to not be expired as 'old' when they should be). Currently we have no real protection against these (admittedly limited-scope) attacks. Worse, without a reliable realtime clock, we're always fetching our system time from an unauthenticated source. We have had problems with this in practice - e.g., captive portals that intercept HTTP traffic and return their own headers, sometimes with wildly bogus times in them.

We should instead use the following sources:
  1. tlsdate(1) pointed at
  2. A file stored in the stateful partition with the last known good time (from tlsdate) and the last known rtc time.
    • The file system itself has a "last modified" timestamp that might be more straightforward -- jrbarnette
    • This can get tricky though as the mtime field could be updated for various reasons, some of which are susceptible to the aforementioned attack vectors.  Having dedicated storage would be clear as to its meaning, and the micro optimization here probably doesn't gain much in trading off to complexity.
  3. The machine's rtc

Of these sources, we can always consider 1 authoritative if we do get a response, since we'll be doing SSL validation of the whole connection. For situations where we can't do 1 (for example, if we're behind a captive portal that is blocking our TLS connection), we can look at the machine's rtc and compare it against the file mentioned in 2; if the rtc is ahead of the file by some reasonable offset, we can trust the rtc and update the file. If the machine's rtc isn't ahead of the file (or is too far ahead), we probably have no reliable time sources and will have to fall back on setting our time to the last good time stored in the file.

There are a couple of other sources we could try: on machines with a 3g modem, we might have access to the network time, but doing this is likely to be fraught with peril. There might be other pieces of hardware on the system (the TPM?) that have a notion of what time it is.

There are some upsides of this crazy plan:

  1. Drop htpdate, which is about 850 lines of code that parses HTTP, in exchange for tlsdate, which is 500 lines of code that uses OpenSSL very carefully;
  2. Works on systems without rtcs

An open question is whether we should accept tlsdate updates if TLS authentication of the connection fails. If we accept unauthenticated updates, we're vulnerable to clock manipulation in the interval (build time, build time + max offset), with max offset most likely being on the order of a few years, but we have a decent chance of getting an at-least-roughly-accurate time. If we reject unauthenticated updates, we may find ourselves in a situation where we have no reliable notion of the time at all, which would force us to default to the build time. Worse, having no reliable time might make it impossible to authenticate to a given captive portal. For a better UX, it's probably wise to just accept the untrusted time, clamping it to the boundaries as needed, since otherwise users might find themselves mysteriously unable to authenticate to a captive portal.