TPMs, event logs, fine-grained measurements and avoiding fragility in remote-attestation
Trusted Platform Modules are fairly unintelligent devices. They can do some crypto, but they don't have any ability to directly monitor the state of the system they're attached to. This is worked around by having each stage of the boot process "measure" state into registers (Platform Configuration Registers, or PCRs) in the TPM by taking the SHA1 of the next boot component and performing an extend operation. Extend works like this:New PCR value = SHA1(current value||new hash)
ie, the TPM takes the current contents of the PCR (a 20-byte register), concatenates the new SHA1 to the end of that in order to obtain a 40-byte value, takes the SHA1 of this 40-byte value to obtain a 20-byte hash and sets the PCR value to this. This has a couple of interesting properties:
- You can't directly modify the contents of the PCR. In order to obtain a specific value, you need to perform the same set of writes in the same order. If you replace the trusted bootloader with an untrusted one that runs arbitrary code, you can't rewrite the PCR to cover up that fact
- The PCR value is predictable and can be reconstructed by replaying the same series of operations
Thankfully we have more than just the final PCR date. The firmware provides an interface to log each extend operation, and you can read the event log in /sys/kernel/security/tpm0/binary_bios_measurements. You can pull information out of that log and use it to reconstruct the writes the firmware made. Merge those with the writes you performed and you should be able to reconstruct the final TPM state. Hurrah!
The problem is that a lot of what you want to measure into the TPM may vary between machines or change in response to configuration changes or system updates. If you measure every module that grub loads, and if grub changes the order that it loads modules in, you also need to update your calculations of the end result. Thankfully there's a way around this - rather than making policy decisions based on the final TPM value, just use the final TPM value to ensure that the log is valid. If you extract each hash value from the log and simulate an extend operation, you should end up with the same value as is present in the TPM. If so, you know that the log is valid. At that point you can examine individual log entries without having to care about the order that they occurred in, which makes writing your policy significantly easier.
But there's another source of fragility. Imagine that you're measuring every command executed by grub (as is the case in the CoreOS grub). You want to ensure that no inappropriate commands have been run (such as ones that would allow you to modify the loaded kernel after it's been measured), but you also want to permit certain variations - for instance, you might have a primary root filesystem and a fallback root filesystem, and you're ok with either being passed as a kernel argument. One approach would be to write two lines of policy, but there's an even more flexible approach. If the bootloader logs the entire command into the event log, when replaying the log we can verify that the event description hashes to the value that was passed to the TPM. If it does, rather than testing against an explicit hash value, we can examine the string itself. If the event description matches a regular expression provided by the policy then we're good.
This approach makes it possible to write TPM policies that are resistant to changes in ordering and permit fine-grained definition of acceptable values, and which can cleanly separate out local policy, generated policy values and values that are provided by the firmware. The split between machine-specific policy and OS policy allows for the static machine-specific policy to be merged with OS-provided policy, making remote attestation viable even over automated system upgrades.
We've integrated an implementation of this kind of policy into the TPM support code we'd like to integrate into Kubernetes, and CoreOS will soon be generating known-good hashes at image build time. The combination of these means that people using Distributed Trusted Computing under Tectonic will be able to validate the state of their systems with nothing more than a minimal machine-specific policy description.
The support code for all of this should also start making it into other distributions in the near future (the grub code is already in Fedora 24), so with luck we can define a cross-distribution policy format and make it straightforward to handle this in a consistent way even in hetrogenous operating system environments. Remote attestation is a powerful tool for ensuring that your systems are in a valid state, but the difficulty of policy management has been a significant factor in making it difficult for people to deploy in their data centres. Making it easier for people to shield themselves against low-level boot attacks is a big step forward in improving the security of distributed workloads and makes bare-metal hosting a much more viable proposition.
comments