tpm.dev.tutorials/Intro

Introduction to TPMs

Trusted Platform Modules (TPMs) are a large and complex topic, made all the more difficult to explain by the intricate relationships between the relevant concepts. This is an attempt at a simple explanation -- much simpler than reading hundreds of pages of documents, but then too, too light on detail to be immediately useful.

So what is a TPM? Well, it's a cryptographic co-processor with special features to enable "root of trust measurement" (RTM), remote attestation of system state, unlocking of local resources that are kept encrypted (e.g., filesystems), and more. A TPM can do those things, and it can do it with rich authentication and authorization policies.

The standards development organization that publishes TPM specifications is the Trusted Computing Group (TCG).

Typically a TPM is a hardware module, a chip, though there are firmware, virtual, and simulated TPMs as well, all implemented in software.

To simplify things we'll consider only TPM 2.0.

Other parts of this tutorial may cover specific concepts in much more detail.

Goals

The goal of this introductory material is to help readers new to TPMs to understand them well enough to approach the subjects of:

• attestation
• secure boot

and to think about the sorts of things that one can do with TPMs in general, which include:

• device on-boarding
• ascertaining the state of a device (e.g., has it executed only trusted code)
• unlocking of devices using TPM-based authentication and authorization policies (e.g., unlocking a laptop on boot multiple factors such as biometrics, smartcards, passwords, time of day, even interaction with remote services)
• using a TPM as a source of entropy for a running OS

NOTE: At this time this introduction is very much a layman's introduction, and only an introduction. Readers seeking to do software development using TPMs will want to make use of TCG specifications and other resources.

Glossary

For a glossary, see section 4 of TCG TPM 2.0 Library part 1: Architecture.

Core Concepts

Some core concepts in the world of TPMs:

NOTE: We will not cover all of these here.

• cryptography
• hash extension
• cryptographic object naming
• platform configuration registers (PCRs)
• immutability of object public areas
• key hierarchies
• key wrapping
• restricted cryptographic keys
• limited resources
• sessions and authorization
• other object types, mainly non-volatile (NV) indexes
• attestation

We'll assume reader familiarity with the basics of cryptography -- the basics of cryptographic primitives as interfaces, but not their internals. E.g., hash functions, symmetric encryption, asymmetric encryption, and digital signatures.

Authorization is the most important aspect of a TPM, since that's ultimately what it exists for: to authorize a system or application to perform certain duties when all the desired conditions allow for it.

TPMs have a very rich set of options for authorization. It's not just policies, but also cryptographic object names used with restricted keys to allow access only to applications that also have other access.

Where to start? Let's start with hash extension.

Hash Extension

Hash extension is just appending some data to a current digest-sized value, hashing that, and then calling the output the new current value:

  v_0 = 0         # size-of-digest-output zero bits
  v_1 = Extend(v_0, e_0)
      = H(v_0 || e_0)
  v_2 = Extend(v_1, e_1)
      = H(v_1 || e_1)
      = H(H(v_0 || e_0) || e_1)
  v_3 = Extend(v_2, e_2)
      = H(v_2 || e_2)
      = H(H(v_1 || e_1) || e_2)
      = H(H(H(v_0 || e_0) || e_1) || e_2)
  ..
  v_n = Extend(v_n-1, e_n-1)
      = H(v_n-1 || e_n-1)
      = H(H(v_n-2 || e_n-2) || e_n-1)
      = H(H(...) || e_n-1)

where H() is a cryptographic hash function.

Each extension value can be arbitrarily large, and the TPM will use the traditional Init/Update/Final approach to making digest computation online.

Note that H(e0 || e1 || e2) != Extend(Extend(Extend(0, e0), e1), e2). Hash extension makes "message" boundaries strong.

Hash extension is most of what a PCR is, but hash extension is used in other TPM concepts besides PCRs, such as policy naming.

Coping with Severe Resource Limits Using Digests and Hash Extension

Hardware TPMs are extremely limited in memory and non-volatile memory capacity. As a result they cannot hold large entities.

A common theme in TPMs is the use of digests, and hash extension digests in particular, as a stand-in for large entities that might not fit at once on the TPM.

TPMs use digests as stand-ins for large entities of various types:

• eventlogs
• policies
• auditing

We'll discuss at least two of those: event logs, and policies.

Platform Configuration Registers (PCRs)

A PCR, then, is just a hash extension output. The only operations on PCRs are: read, extend, and reset. All richness of semantics of PCRs come from how they are used:

• what the governing TCG platform specification says about them
• what they are extended with and by what code (in what locality)
• what purposes they are read for
  • attestation
  • authorization

Note that a PCR value by itself is devoid of meaning. To add meaning one must have access to the list of discrete values extended into the PCR, as well as the order in which they were extended into the PCR. And one must know the meaning of each such value.

Eventlogs

Any TPM-using platform has to provide a way to keep a log of PCR hash extension values. Such a log is known as the "eventlog".

The TPM itself cannot hold this log -- the TPM is too resource-constrained.

Root of Trust Measurements (RTM)

When a computer and its TPM start up, most PCRs are set to all-zeros, and then the computer's boot firmware performs a core root of trust measurement (CRTM) to "measure" (i.e., hash) the the next boot stage and extend it into an agreed-upon PCR. The entire boot process should, ideally, perform RTMs. At the end of the boot process some set of PCRs should reflect the totality of the code path taken to complete booting.

Some PCRs are used to measure the BIOS, others to measure option ROMs, and others to measure the operating system. Each platform has a specification for which PCRs are used or reserved for what purposes. In principle one could measure the entirety of an operating system and all the code that is installed on the system.

RTM can be used to ensure that only known-trusted code is executed, and that important resources are not unlocked unless the state of the system when they are needed is "has only executed trusted code to get here".

Note that some PCRs are left to be used by "applications".

Some terms:

• core RTM (CRTM) -- initial measurements performed upon power-on
• static RTM (SRTM) -- subsequent-to-CRTM measurements of next boot stages
• dynamic RTM (DRTM) -- measurements involved in rebooting

Resource unlocking can be done by creating objects tied to a set of PCRs such that they must each have specific values for the TPM to be willing to unlock (unseal) the object.

The PCR Extension Eventlog

On the "PC platform" (which includes x64 servers) the BIOS keeps a log of all the PCR extensions it has performed. The OS should keep its own log of extensions it performs of PCRs reserved to the OS. Each application has to keep a log of the extensions of the PCRs allocated to it. Again, the TPM itself cannot do this.

The eventlog documents how each PCR evolved to their current state, whatever it might be. Since PCR extension values are typically digests, the eventlog is very dry, but it can still be used to evaluate whether the current PCR values represent a trusted state. For example, one might have a database of known-good and known-bad firmware/ROM digests, then one can check that only known-good ones appear in the eventlog and that reproducing the hash extensions described by the eventlog produces the same PCR values as one can read, and if so it follows that the system has only executed trusted code to arrive at the state identified by the PCRs.

Note though that PCRs and RTM are not enough on their own to keep a system from executing untrusted code. A system can be configured to allow execution of arbitrary code at some point (e.g., download and execute) and to not extend PCRs accordingly, in which case the execution of untrusted code will not be reflected in any RTM.

Tickets

Tickets are yet another device for coping with TPMs having limited resources. Interaction with TPMs is via request/response commands, and tickets are largely about making TPMs stateless between related commands.

To avoid having to re-perform various operations -or remember having performed them- between command invocations, a TPM can produce a "ticket" that consists of an HMAC over a TPM-generated assertion, keyed by a key known only to the TPM, and return it to the caller who can then present it in a subsequent command related to the first.

For example, when signing data the TPM will first digest the data to sign over several commands and generate a ticket saying it did produce that digest, then later it can sign the digest after validating the ticket that it produced.

Another example is a policy ticket, which allows one to avoid having to re-authenticate (e.g., with smartcard, biometrics, passwords) on every command for small window of time.

When would a user be authenticated? Well, typically at boot time, or maybe at wake from sleep/hibernate time. A laptop could be configured to require a user to authenticate with biometrics and possibly a password or a smartcard. Note that such policies are not required by the specifications, but rather something that one can choose to use.

There are five types of tickets. We won't cover them here. Readers who end up needing to know about them can look at section 11.4.6.3 of TCG TPM 2.0 Library, part 1: Architecture.

Cryptographic Object Naming

TPMs support a variety of types of objects. Objects generally have pointer-like "handles" that they are often used in the TPM APIs. But more importantly, objects have cryptographically-secure names that are used in many cases.

The cryptographically-secure name of an object is the hash of the object's "public area".

The public area of, say, an asymmetric key is a data structure that includes the public key (corresponding to the private key), and various attributes of the key (e.g., its algorithm, but also whether it is bound to the TPM where it resides, or to its key hierarchy), unseal policy, and access policy.

This concept is extremely important. Because object names are the outputs of cryptographically strong digest (hash) functions, they are resistant to collision attacks, first pre-image attacks, and second pre-image attacks -- as strong as the hash algorithm used anyways. Which means that object names cannot be forged easily, which means that they can be used in context where a peer needs certain guarantees, or to defeat active attacks.

Immutability of Object Public Areas

Because the name of an object is a digest of its public area, the public area cannot be changed after creating it. One can delete and then recreate an object in order to "change" its public area, but this necessarily yields a new name.

Cryptographic Object Naming as a Binding

This section comes too soon, since it relates to attestation and restricted keys. Still, it may be useful to illustrate cryptographic object naming with one particularly important use of it.

A pair of functions, TPM2_MakeCredential() and TPM2_ActivateCredential(), illustrate the use of cryptographic object naming as a binding or a sort of authorization function.

TPM2_MakeCredential() can be used to encrypt a datum (a "credential") to a target TPM such that the target will only be willing to decrypt it if and only if the application calling TPM2_ActivateCredential() to decrypt that credential has access to some key named by the sender, and that name is a cryptographic name that the sender can and must compute for itself.

The semantics of these two functions can be used to defeat a cut-and-paste attack in attestation protocols.

Key Hierarchies

TPMs have multiple key hierarchies, each with zero, one or more primary keys, each with zero, one, or more children keys:

                seed
                /|\
               / | \
              v  v  v
     primary key (asymmetric encryption)
                /|\
               / | \
              v  v  v
       secondary keys (of any kind)
                /|\
               / | \
              v  v  v
                ...

Keys that have no parent are primary keys.

There are four built-in hierarchies:

• platform hierarchy
• endorsement hierarchy
• storage hierarchy
• null hierarchy

of which only the endorsement and storage hierarchies will be of interest to most readers.

The endorsement hierarchy is used to authenticate (when needed) that a TPM is a legitimate TPM. The primary endorsement key is known as the EK (endorsement key). Hardware TPMs come with a certificate for the EK issued by the TPM's manufacturer. This EK certificate ("EKcert") can be used to authenticate the TPM's legitimacy. The EK's public key ("EKpub") can be used to uniquely identify a TPM, and possibly link to the platform's, and even the platform's user(s)' identities.

The TPM2_CreatePrimary() command creates primary key objects deterministically from the hierarchy's seed and the "template" used to create the key (which includes a "unique" area that provides "entropy" to the key derivation function).

The TPM2_Create() command creates a ordinary objects.

The TPM2_CreateLoaded() command can also create primary key objects deterministically from the hierarchy's seed and the "template" used to create the key (which includes a "unique" area that provides "entropy" to the key derivation function).

Key Wrapping and Resource Management

Key wrapping is encrypting a secret or private key (key encryotion key, or KEK) such that a specific entity may recover the plain key.

A decrypt-only asymmetric private key can be used to encrypt secrets to the TPM on which that private key resides.

As well as wrapping secrets by encryption to public keys, TPMs also use wrapping in a symmetric key known only to the TPM for the purpose of saving keys off the TPM.

This is used for resource management: since hardware TPMs have very limited resources, objects need to created or loaded, used, then saved off-TPM to make room for other objects to be loaded (unless they are not to be used again, then saving them is pointless). Only a TPM that saved an object can load it again, but some objects can be exported to other TPMs by encrypting them to their destination TPMs' EKpubs.

With a resource manager and access broker, a TPM can appear to have infinite resources.

Controlling Exportability of Keys

A key that is fixedTPM cannot leave the TPM in cleartext. It can be saved off the TPM it resides in, but only that TPM can load it again.

A key that is fixedParent cannot be moved from one part of a key hierarchy to another -- it cannot be "re-parented". Though if its parent is neither fixedParent nor fixedTPM then the parent and its descendants can be moved as a group to some other TPM.

Key hierarchies are an important TPM topic that we're not really addresing in this intro.

Persistence

In a TPM, key objects are, by default, transient, meaning the TPM will forget them if restarted. Still, they can be saved (encrypted in a secret key only the TPM knows) and later reloaded.

Transient objects can be made persistent, but because hardware TPMs have very little non-volatile memory, few keys should be made persistent. Instead you can save them (wrapped to a TPM-only KEK) and reload them as needed.

Because primary keys (for any hierarchy other than the null hierarchy) are derived deterministically from a built-in and protected seed, and from a template, they are persistent even when not moved to NV storage and even when not saved as long as the hierarchy's seed is not reset.

(Resetting the endorsement hierarchy seed is a very dramatic action, as it changes the EK/EKpub and renders any provisioned EKcert useless. Resetting the storage hierarchy seed is much less dramatic. The NULL hierarchy is reset every time the TPM resets.)

PCRs always persist, but they get reset on restart.

NV indexes always persist. (But in disorderly resets/shutdowns a hybrid NV index may not be sync'ed to NV.)

Non-Volatile (NV) Indexes

TPMs also have a special kind of non-volatile object: NV indexes.

NOTE: NV indexes are not "objects" in the sense that the TCG's specifications mean. TCG's definition of "object" is

key or data that has a public portion and, optionally, a sensitive portion; and which is a member of a hierarchy

NV indexes come in multiple flavors for various uses:

• store public data (e.g., an NV index is used to store the EKcert)
• emulate PCRs
• monotonic counters
• fields of write-once bits (bitfields) (for, e.g., revocation)
• ...

NV indexes can be used standalone, and/or in connection with policies, to enrich application TPM semantics.

Authentication and Authorization

TPMs have multiple ways to authenticate users/entities:

• plain passwords (legacy)
• HMAC based on secret keys or passwords
• public key signed attestations of identification by biometric authentication devices

TPMs have two ways to authorize access to various objects:

• plain passwords (legacy)
• HMAC proof of possession of a secret key or password
• arbitrarily complex policies that can make reference to:
  • PCR state
  • NV index state
  • time of day
  • authentication state
  • etc.

Policies

A policy consists of a sequence of "commands" that each asserts something of interest.

Policies are particularly interesting because they are cryptographically named using hash extension with the sequence of "commands" that make up a policy. Therefore no matter how complex and large a policy is, it is always "compressed" to a hash digest.

It is the responsibility of the application that will attempt to use a policy-protected resource to know what the policy's definition is and restate it to the TPM when it goes to make use of that resource. The TPM will evaluate the policy and, at the end, check that its digest matches that of the policy-protected resource. Thus, and because policy digests are small and fixed-sized, they can be arbitrarily more complex than a TPM's limited resources would otherwise allow.

All the policy commands that are to be evaluated successfully to grant access have to be known to the entity that wants that access. Of course, that entity will have to satisfy -at access time- the conditions expressed by the relevant policy. And that entity has to know the policy because the TPM knows only a digest of it.

Policy Construction

Construction of a policy consists of computing it by hash extending an initial all-zeroes value with the commands that make up the policy.

Policy Evaluation

Evaluation of a policy consists of issuing those same commands to the TPM in a session, with those commands either evaluated immediately or deferred to the time of execution of the to-be-authorized command, but the TPM computes the same hash extension as it goes. Once all policy commands being evaluated have succeeded, the resulting hash extension value is compared to the policy that protects the resource(s) being used by the to-be-authorized command, and if it matches, then the command is allowed, otherwise it is not.

Indirect Policies

Because an object's policy is part of its name, that policy cannot be changed after creation. An indirect policy command allows for a policy to change over time without having to recreate the authorized object.

Compound Policies

Policies consist of a conjunction (logical-AND) of assertions that must be true at evaluation time.

However, there is a special policy command that allows for alternation (logical-OR). This policy command has a number of alternative policy digests. At evaluation time, one of the alternation digests must match the extension value for the policy commands up to (but excluding) the logical-OR policy command. At evaluation time the caller must have picked one of the alternatives and executed the commands that make it up.

(Recall that the application has to know the definition of the policy because the TPM knows only the policy's digest.)

Rich Policy Semantics

With all these features, and with all the flexibility allowed by NV indexes, policies can be used to:

• require that N-of-M users authenticate
• require multi-factor authentication (password, biometric, smartcard)
• enforce bank vault-like time of day restrictions
• check revocation (using NV index bit-field objects)
• check system RTM state (PCRs)
• distinguish user roles

Sessions

A session is an object (meaning, among other things, that it can be loaded and unloaded as needed) that represents the current policy construction or evaluation hash extension digest (the policyDigest), and the objects that have been granted access.

Restricted Cryptographic Keys

Cryptographic keys can either be unrestricted or restricted.

An unrestricted signing key can be used to sign arbitrary content.

An unrestricted decryption key can be used to decrypt arbitrary ciphertexts encrypted to that key's public key.

NOTE WELL: The endorsement key (EK) is a restricted key.

Restricted Signing Keys

A restricted signing key can be used to sign only TPM-generated content as part of specific TPM restricted signing commands. Such content always begins with a magic byte sequence. Conversely, the TPM refuses to sign externally generated content that starts with that magic byte sequence. See the TPM2_Certify(), TPM2_Quote(), TPM2_CertifyCreation(), TPM2_GetSessionAuditDigest(), and TPM2_GetCommandAuditDigest() TPM commands.

There is also a notion of signing keys that can only be used to sign PKIX certificates using TPM2_CertifyX509().

Restricted Decryption Keys

NOTE WELL: The endorsement key (EK) is a restricted key.

A restricted decryption key can only be used to decrypt ciphertexts whose plaintexts have a certain structure. In particular these are used for TPM2_MakeCredential() / TPM2_ActivateCredential() to allow the TPM-using application to get the plaintext if and only if (IFF) the plaintext cryptographically names an object that the application has access to. This is used to communicate secrets ("credentials") to TPMs.

Another operation that a restricted decryption key can perform is TPM2_Import(), which decrypts a key wrapped to the given decrypt-only key and outputs a file that can be loaded with TPM2_Load(). The wrapped key payload given to TPM2_Import() too has a particular structure and is produced by a remote peer using TPM2_Duplicate().

To recap, a restricted decryption key can only be used to:

• "activate credentials" (made with TPM2_MakeCredential())

• receive wrapped keys sent by a peer (made with TPM2_Duplicate())

Attestation

Attestation is the process of demonstrating that a system's current state is "trusted", or the truthfulness of some set of assertions.

Often a system gets something in exchange for attesting to its current state. E.g., keys for unlocking filesystems, or device credentials.

As you can see in our tutorial on attestation, many TPM concepts can be used to great effect:

• using PCRs to attest to system state
• using policies and sealed-to-PCRs objects to attest to authentication events on the system
• using restricted keys and cryptographic object naming to authenticate a TPM and bind it to its host
• delivering key material to authenticated systems via their TPMs
• unlocking resources following successful attestation
• authorization of devices onto a network
• etc.

Other Resources

A Practical Guide to TPM 2.0 is an excellent book that informed much of this tutorial.

Nokia has a TPM course.

The TCG has a number of members-only tutorials, but it seems that it is possible to be invited to be a non-fee paying member.

Core TCG TPM specs:

• TCG TPM 2.0 Library part 1: Architecture.
• TCG TPM 2.0 Library part 2: Structures.
• TCG TPM 2.0 Library part 3: Commands, section 12.
• TCG TPM 2.0 Library part 3: Commands Code, section 12.