Most modern PCs have a Trusted Platform Module (TPM) and firmware that, together, support something called Trusted Boot. In Trusted Boot, each component in the boot chain generates a series of measurements of next component of the boot process and relevant configuration. These measurements are pushed to the TPM where they’re combined with the existing values stored in a series of Platform Configuration Registers (PCRs) in such a way that the final PCR value depends on both the value and the order of the measurements it’s given. If any measurements change, the final PCR value changes.
Windows takes advantage of this with its Bitlocker disk encryption technology. The disk encryption key is stored in the TPM along with a policy that tells it to release it only if a specific set of PCR values is correct. By default, the TPM will release the encryption key automatically if the PCR values match and the system will just transparently boot. If someone tampers with the boot process or configuration, the PCR values will no longer match and boot will halt to allow the user to provide the disk key in some other way.
Unfortunately the TPM keeps no record of how it got to a specific state. If the PCR values don’t match, that’s all we know – the TPM is unable to tell us what changed to result in this breakage. Fortunately, the system firmware maintains an event log as we go along. Each measurement that’s pushed to the TPM is accompanied by a new entry in the event log, containing not only the hash that was pushed to the TPM but also metadata that tells us what was measured and why. Since the algorithm the TPM uses to calculate the hash values is known, we can replay the same values from the event log and verify that we end up with the same final value that’s in the TPM. We can then examine the event log to see what changed.
Unfortunately, the event log is stored in unprotected system RAM. In order to be able to trust it we need to compare the values in the event log (which can be tampered with) with the values in the TPM (which are much harder to tamper with). Unfortunately if someone has tampered with the event log then they could also have tampered with the bits of the OS that are doing that comparison. Put simply, if the machine is in a potentially untrustworthy state, we can’t trust that machine to tell us anything about itself.
This is solved using a procedure called Remote Attestation. The TPM can be asked to provide a digital signature of the PCR values, and this can be passed to a remote system along with the event log. That remote system can then examine the event log, make sure it corresponds to the signed PCR values and make a security decision based on the contents of the event log rather than just on the final PCR values. This makes the system significantly more flexible and aids diagnostics. Unfortunately, it also means you need a remote server and an internet connection and then some way for that remote server to tell you whether it thinks your system is trustworthy and also you need some way to believe that the remote server is trustworthy and all of this is well not ideal if you’re not an enterprise.
Last week I gave a talk at linux.conf.au on one way around this. Basically, remote attestation places no constraints on the network protocol in use – while the implementations that exist all do this over IP, there’s no requirement for them to do so. So I wrote an implementation that runs over Bluetooth, in theory allowing you to use your phone to serve as the remote agent. If you trust your phone, you can use it as a tool for determining if you should trust your laptop.
I’ve pushed some code that demos this. The current implementation does nothing other than tell you whether UEFI Secure Boot was enabled or not, and it’s also not currently running on a phone. The phone bit of this is pretty straightforward to fix, but the rest is somewhat harder.
The big issue we face is that we frequently don’t know what event log values we should be seeing. The first few values are produced by the system firmware and there’s no standardised way to publish the expected values. The Linux Vendor Firmware Service has support for publishing these values, so for some systems we can get hold of this. But then you get to measurements of your bootloader and kernel, and those change every time you do an update. Ideally we’d have tooling for Linux distributions to publish known good values for each package version and for that to be common across distributions. This would allow tools to download metadata and verify that measurements correspond to legitimate builds from the distribution in question.
This does still leave the problem of the initramfs. Since initramfs files are usually generated locally, and depend on the locally installed versions of tools at the point they’re built, we end up with no good way to precalculate those values. I proposed a possible solution to this a while back, but have done absolutely nothing to help make that happen. I suck. The right way to do this may actually just be to turn initramfs images into pre-built artifacts and figure out the config at runtime (dracut actually supports a bunch of this already), so I’m going to spend a while playing with that.
If we can pull these pieces together then we can get to a place where you can boot your laptop and then, before typing any authentication details, have your phone compare each component in the boot process to expected values. Assistance in all of this extremely gratefully received.