This document discusses the security architecture of systems where PSA API functions might receive arguments that are in memory that is shared with an untrusted process. On such systems, the untrusted process might access a shared memory buffer while the cryptography library is using it, and thus cause unexpected behavior in the cryptography code.
### Core assumptions
We assume the following scope limitations:
* Only PSA Crypto API functions are in scope (including Mbed TLS extensions to the official API specification). Legacy crypto, X.509, TLS, or any other function which is not called `psa_xxx` is out of scope.
* We only consider [input buffers](https://arm-software.github.io/psa-api/crypto/1.1/overview/conventions.html#input-buffer-sizes) and [output buffers](https://arm-software.github.io/psa-api/crypto/1.1/overview/conventions.html#output-buffer-sizes). Any other data is assumed to be in non-shared memory.
We consider a system that has memory separation between partitions: a partition can't access another partition's memory directly. Partitions are meant to be isolated from each other: a partition may only affect the integrity of another partition via well-defined system interfaces. For example, this can be a Unix/POSIX-like system that isolates processes, or isolation between the secure world and the non-secure world relying on a mechanism such as TrustZone, or isolation between secure-world applications on such a system.
More precisely, we consider such a system where our PSA Crypto implementation is running inside one partition, called the **crypto service**. The crypto service receives remote procedure calls (RPC) from other partitions, validates their arguments (e.g. validation of key identifier ownership), and calls a PSA Crypto API function. This document is concerned with environments where the arguments passed to a PSA Crypto API function may be in shared memory (as opposed to environments where the inputs are always copied into memory that is solely accessible by the crypto service before calling the API function, and likewise with output buffers after the function returns).
When the data is accessible to another partition, there is a risk that this other partition will access it while the crypto implementation is working. Although this could be prevented by suspending the whole system while crypto is working, such a limitation is rarely desirable and most systems don't offer a way to do it. (Even systems that have absolute thread priorities, and where crypto has a higher priority than any untrusted partition, may be vulnerable due to having multiple cores or asynchronous data transfers with peripherals.)
The crypto service must guarantee that it behaves as if the rest of the world was suspended while it is executed. A behavior that is only possible if an untrusted entity accesses a buffer while the crypto service is processing the data is a security violation.
The behavior of RPC is defined for in terms of values of inputs and outputs. This models an ideal world where the content of input and output buffers is not accessible outside the crypto service while it is processing an RPC. It is a security violation if the crypto service behaves in a way that cannot be achieved by setting the inputs before the RPC call, and reading the outputs after the RPC call is finished.
Vulnerability example (parsing): suppose the input contains data with a type-length-value or length-value encoding (for example, importing an RSA key). The crypto code reads the length field and checks that it fits within the buffer. (This could be the length of the overall data, or the length of an embedded field) Later, the crypto code reads the length again and uses it without validation. A malicious client can modify the length field in the shared memory between the two reads and thus cause a buffer overread on the second read.
Vulnerability example (dual processing): consider an RPC to perform authenticated encryption, using a mechanism with an encrypt-and-MAC structure. The authenticated encryption implementation separately calculates the ciphertext and the MAC from the plaintext. A client sets the plaintext input to `"PPPP"`, then starts the RPC call, then changes the input buffer to `"QQQQ"` while the crypto service is working.
* Any of `enc("PPPP")+mac("PPPP")`, `enc("PPQQ")+mac("PPQQ")` or `enc("QQQQ")+mac("QQQQ")` are valid outputs: they are outputs that can be produced by this authenticated encryption RPC.
* If the authenticated encryption calculates the ciphertext before the client changes the output buffer and calculates the MAC after that change, reading the input buffer again each time, the output will be `enc("PPPP")+mac("QQQQ")`. There is no input that can lead to this output, hence this behavior violates the security guarantees of the crypto service.
3. The crypto code reads the intermediate data back and continues the calculation, leading to an outcome that would not be possible if the intermediate data had not been modified.
Vulnerability example: suppose that an RSA signature function works by formatting the data in place in the output buffer, then applying the RSA private-key operation in place. (This is how `mbedtls_rsa_pkcs1_sign` works.) A malicious client may write badly formatted data into the buffer, so that the private-key operation is not a valid signature (e.g. it could be a decryption), violating the RSA key's usage policy.
Vulnerability example with chained calls: we consider the same RSA signature operation as before. In this example, we additionally assume that the data to sign comes from an attestation application which signs some data on behalf of a final client: the key and the data to sign are under the attestation application's control, and the final client must not be able to obtain arbitrary signatures. The final client shares an output buffer for the signature with the attestation application, and the attestation application re-shares this buffer with the crypto service. A malicious final client can modify the intermediate data and thus sign arbitrary data.
#### Write-write disclosure
If an output argument is in shared memory, there is a risk of a **write-write disclosure**:
1. The crypto code writes some intermediate data into the output buffer. This intermediate data must remain confidential.
2. The client (or client's client) reads the intermediate data.
3. The crypto code overwrites the intermediate data.
Vulnerability example with chained calls (temporary exposure): an application encrypts some data, and lets its clients store the ciphertext. Clients may not have access to the plaintext. To save memory, when it calls the crypto service, it passes an output buffer that is in the final client's memory. Suppose the encryption mechanism works by copying its input to the output buffer then encrypting in place (for example, to simplify considerations related to overlap, or because the implementation relies on a low-level API that works in place). In this scenario, the plaintext is exposed to the final client while the encryption in progress, which violates the confidentiality of the plaintext.
Vulnerability example with chained calls (backtrack): we consider a provisioning application that provides a data encryption service on behalf of multiple clients, using a single shared key. Clients are not allowed to access each other's data. The provisioning application isolates clients by including the client identity in the associated data. Suppose that an AEAD decryption function processes the ciphertext incrementally by simultaneously writing the plaintext to the output buffer and calculating the tag. (This is how AEAD decryption usually works.) At the end, if the tag is wrong, the decryption function wipes the output buffer. Assume that the output buffer for the plaintext is shared from the client to the provisioning application, which re-shares it with the crypto service. A malicious client can read another client (the victim)'s encrypted data by passing the ciphertext to the provisioning application, which will attempt to decrypt it with associated data identifying the requesting client. Although the operation will fail beacuse the tag is wrong, the malicious client still reads the victim plaintext.
If a function both has an input argument and an output argument in shared memory, and processes its input incrementally to emit output incrementally, the following sequence of events is possible:
1. The crypto code processes part of the input and writes the corresponding part of the output.
2. The client reads the early output and uses that to calculate the next part of the input.
3. The crypto code processes the rest of the input.
There are cryptographic mechanisms for which this breaks security properties. An example is [CBC encryption](https://link.springer.com/content/pdf/10.1007/3-540-45708-9_2.pdf): if the client can choose the content of a plaintext block after seeing the immediately preceding ciphertext block, this gives the client a decryption oracle. This is a security violation if the key policy only allowed the client to encrypt, not to decrypt.
TODO: is this a risk we want to take into account? Although this extends the possible behaviors of the one-shot interface, the client can do the same thing legitimately with the multipart interface.
In this section, we briefly discuss generic countermeasures.
#### Copying
Copying is a valid countermeasure. It is conceptually simple. However, it is often unattractive because it requires additional memory and time.
Note that although copying is very easy to write into a program, there is a risk that a compiler (especially with whole-program optimization) may optimize the copy away, if it does not understand that copies between shared memory and non-shared memory are semantically meaningful.
Example: the PSA Firmware Framework 1.0 forbids shared memory between partitions. This restriction is lifted in version 1.1 due to concerns over RAM usage.
* Never write to the output twice at the same index.
* This rule can usefully be relaxed in many circumstances. It is ok to write data that is independent of the inputs (and not otherwise confidential), then overwrite it. For example, it is ok to zero the output buffer before starting to process the input.
Example: these are the rules that a GlobalPlatform TEE Trusted Application (application running on the secure side of TrustZone on Cortex-A) must follow.
The [PSA Crypto API specification](https://arm-software.github.io/psa-api/crypto/1.1/overview/conventions.html#stability-of-parameters) puts the responsibility for protection on the implementation of the PSA Crypto API, i.e. (3) or (4).
> In an environment with multiple threads or with shared memory, the implementation carefully accesses non-overlapping buffer parameters in order to prevent any security risk resulting from the content of the buffer being modified or observed during the execution of the function. (...)
In Mbed TLS 2.x and 3.x up to and including 3.5.0, there is no defense against buffers in shared memory. The responsibility shifts to (1) or (2), but this is not documented.
In the remainder of this chapter, we will discuss how to implement this high-level requirement where it belongs: inside the implementation of the PSA Crypto API. Note that this allows two possible levels: in the dispatch layer (independently of the implementation of each mechanism) or in the driver (specific to each implementation).
#### Protection in the dispatch layer
The dispatch layer has no control over how the driver layer will access buffers. Therefore the only possible protection at this layer method is to ensure that drivers have no access to shared memory. This means that any buffer located in shared memory must be copied into or out of a buffer in memory owned by the crypto service (heap or stack). This adds inefficiency, mostly in terms of RAM usage.
For buffers with a small static size limit, this is something we often do for convenience, especially with output buffers. However, as of Mbed TLS 3.5.0, it is not done systematically.
It is ok to skip the copy if it is known for sure that a buffer is not in shared memory. However, the location of the buffer is not under the control of Mbed TLS. This means skipping the copy would have to be a compile-time or run-time option which has to be set by the application using Mbed TLS. This is both an additional maintenance cost (more code to analyze, more testing burden), and a residual security risk in case the party who is responsible for setting this option does not set it correctly. As a consequence, Mbed TLS will not offer this configurability unless there is a compelling argument.
#### Protection in the driver layer
Putting the responsibility for protection in the driver layer increases the overall amount of work since there are more driver implementations than dispatch implementations. (This is true even inside Mbed TLS: almost all API functions have multiple underlying implementations, one for each algorithm.) It also increases the risk to the ecosystem since some drivers might not protect correctly. Therefore having drivers be responsible for protection is only a good choice if there is a definite benefit to it, compared to allocating an internal buffer and copying. An expected benefit in some cases is that there are practical protection methods other than copying.
Some cryptographic mechanisms are naturally implemented by processing the input in a single pass, with a low risk of ever reading the same byte twice, and by writing the final output directly into the output buffer. For such mechanism, it is sensible to mandate that drivers respect these rules.
In the next section, we will analyze how susceptible various cryptographic mechanisms are to shared memory vulnerabilities.
### Susceptibility of different mechanisms
#### Operations involving small buffers
For operations involving **small buffers**, the cost of copying is low. For many of those, the risk of not copying is high:
* Any parsing of formatted data has a high risk of [read-read inconsistency](#read-read-inconsistency).
* An internal review shows that for RSA operations, it is natural for an implementation to have a [write-read inconsistency](#write-read-inconsistency) or a [write-write disclosure](#write-write-disclosure).
Note that in this context, a “small buffer” is one with a size limit that is known at compile time, and small enough that copying the data is not prohibitive. For example, an RSA key fits in a small buffer. A hash input is not a small buffer, even if it happens to be only a few bytes long in one particular call.
The following buffers are considered small buffers:
* Any input or output directly related to asymmetric cryptography (signature, encryption/decryption, key exchange, PAKE), including key import and export.
* Note that this does not include inputs or outputs that are not processed by an asymmetric primitives, for example the message input to `psa_sign_message` or `psa_verify_message`.
Message inputs to hash, MAC and key derivation operations are at a low risk of [read-read inconsistency](#read-read-inconsistency) because they are unformatted data, and for all specified algorithms, it is natural to process the input one byte at a time.
**Design decision: require symmetric cryptography drivers to read their input without a risk of read-read inconsistency**.
TODO: what about IV/nonce inputs? They are typically small, but don't necessarily have a static size limit (e.g. GCM recommends a 12-byte nonce, but also allows large nonces).
#### Key derivation outputs
Key derivation typically emits its output as a stream, with no error condition detected after setup other than operational failures (e.g. communication failure with an accelerator) or running out of data to emit (which can easily be checked before emitting any data, since the data size is known in advance).
(Note that this is about raw byte output, not about cooked key derivation, i.e. deriving a structured key, which is considered a [small buffer](#operations-involving-small-buffers).)
**Design decision: require key derivation drivers to emit their output without reading back from the output buffer**.
AEAD encryption and decryption are at risk of [read-read inconsistency](#read-read-inconsistency) if they process the input multiple times, which is natural in a number of cases:
* when encrypting with an encrypt-and-authenticate or authenticate-then-encrypt structure (one read to calculate the authentication tag and another read to encrypt);
* when decrypting with an encrypt-then-authenticate structure (one read to decrypt and one read to calculate the authentication tag);
* with SIV modes (not yet present in the PSA API, but likely to come one day) (one full pass to calculate the IV, then another full pass for the core authenticated encryption);
Cipher and AEAD outputs are at risk of [write-read inconsistency](#write-read-inconsistency) and [write-write disclosure](#write-write-disclosure) if they are implemented by copying the input into the output buffer with `memmove`, then processing the data in place. In particular, this approach makes it easy to fully support overlapping, since `memmove` will take care of overlapping cases correctly, which is otherwise hard to do portably (C99 does not offer an efficient, portable way to check whether two buffers overlap).
**Design decision: the dispatch layer shall allocate an intermediate buffer for cipher and AEAD plaintext/ciphertext inputs and outputs**.
Note that this can be a single buffer for the input and the output if the driver supports in-place operation (which it is supposed to, since it is supposed to support arbitrary overlap, although this is not always the case in Mbed TLS, a [known issue](https://github.com/Mbed-TLS/mbedtls/issues/3266)). A side benefit of doing this intermediate copy is that overlap will be supported.
**Design decision: for now, require AEAD drivers to read the additional data without a risk of read-read inconsistency**. Make a note to revisit this when we start supporting an SIV mode, at which point the dispatch layer shall copy the input for modes that are not known to be low-risk.
For signature algorithms with a hash-and-sign framework, the input to sign/verify-message is passed to a hash, and thus can follow the same rules as [symmetric cryptography inputs with small output](#symmetric-cryptography-inputs-with-small-output). This is also true for `PSA_ALG_RSA_PKCS1V15_SIGN_RAW`, which is the only non-hash-and-sign signature mechanism implemented in Mbed TLS 3.5. This is not true for PureEdDSA (`#PSA_ALG_PURE_EDDSA`), which is not yet implemented: [PureEdDSA signature](https://www.rfc-editor.org/rfc/rfc8032#section-5.1.6) processes the message twice. (However, PureEdDSA verification only processes the message once.)
**Design decision: for now, require sign/verify-message drivers to read their input without a risk of read-read inconsistency**. Make a note to revisit this when we start supporting PureEdDSA, at which point the dispatch layer shall copy the input for algorithms such as PureEdDSA that are not known to be low-risk.
Copy what needs copying. This is broadly straightforward, however there are a few things to consider.
#### Compiler optimization of copies
It is unclear whether the compiler will attempt to optimize away copying operations.
Once the copying code is implemented, it should be evaluated to see whether compiler optimization is a problem. Specifically, for the major compilers and platforms supported by Mbed TLS:
* Build Mbed TLS with the compiler set to a high level of optimization (e.g. `-O2` for `gcc`).
* Inspect the generated code with `objdump` or a similar tool to see if copying operations are preserved.
If copying behaviour is preserved by all major compilers and platforms then assume that compiler optimization is not a problem.
If copying behaviour is optimized away by the compiler, further investigation is needed. Experiment with using the `volatile` keyword to force the compiler not to optimize accesses to the copied buffers.
**Open questions: Will the compiler optimize away copies? If so, can it be prevented from doing so in a portable way?**
#### Copying code
We may either copy buffers on an ad-hoc basis using `memcpy()` in each PSA function, or use a unified set of functions for copying input and output data. The advantages of the latter are obvious:
* Any test hooks need only be added in one place.
* Copying code must only be reviewed for correctness in one place, rather than in all functions where it occurs.
* Copy bypass is simpler as we can just replace these functions with no-ops in a single place.
* Any complexity needed to prevent the compiler optimizing copies away does not have to be duplicated.
On the other hand, the only advantage of ad-hoc copying is slightly greater flexibility.
**Design decision: Create a unified set of functions for copying input and output data.**
#### Copying in multipart APIs
Multipart APIs may follow one of 2 possible approaches for copying of input:
##### 1. Allocate a buffer and copy input on each call to `update()`
This is simple and mirrors the approach for one-shot APIs nicely. However, allocating memory in the middle of a multi-part operation is likely to be bad for performance. Multipart APIs are designed in part for systems that do not have time to perform an operation at once, so introducing poor performance may be a problem here.
**Open question: Does memory allocation in `update()` cause a performance problem? If so, to what extent?**
##### 2. Allocate a buffer at the start of the operation and subdivide calls to `update()`
In this approach, input and output buffers are allocated at the start of the operation that are large enough to hold the expected average call to `update()`. When `update()` is called with larger buffers than these, the PSA API layer makes multiple calls to the driver, chopping the input into chunks of the temporary buffer size and filling the output from the results until the operation is finished.
This would be more complicated than approach (1) and introduces some extra issues. For example, if one of the intermediate calls to the driver's `update()` returns an error, it is not possible for the driver's state to be rolled back to before the first call to `update()`. It is unclear how this could be solved.
However, this approach would reduce memory usage in some cases and prevent memory allocation during an operation. Additionally, since the input and output buffers would be fixed-size it would be possible to allocate them statically, avoiding the need for any dynamic memory allocation at all.
**Design decision: Initially use approach (1) and treat approach (2) as an optimization to be done if necessary.**
This is fairly self-explanatory. Review all functions that use shared memory and ensure that they each copy memory. This is the simplest strategy to implement but is less reliable than automated validation.
Proposed general idea: have tests where the test code calling API functions allocates memory in a certain pool, and code in the library allocates memory in a different pool. Test drivers check that needs-copying arguments are within the library pool, not within the test pool.
Proposed general idea: in test code, “poison” the memory area used by input and output parameters that must be copied. Poisoning means something that prevents accessing memory while it is poisoned. This could be via memory protection (allocate with `mmap` then disable access with `mprotect`), or some kind of poisoning for an analyzer such as MSan or Valgrind.
In the library, the code that does the copying temporarily unpoisons the memory by calling a test hook.
The reason to poison the memory before calling the library, rather than after the copy-in (and symmetrically for output buffers) is so that the test will fail if we forget to copy, or we copy the wrong thing. This would not be the case if we relied on the library's copy function to do the poisoning: that would only validate that the driver code does not access the memory on the condition that the copy is done as expected.
Note: Extra work may be needed when allocating buffers to ensure that each shared buffer lies in its own separate page, allowing its permissions to be set independently.
**Question: Should we try to build memory poisoning validation on existing Mbed TLS tests, or write new tests for this?**
##### Validation with existing tests
It should be possible to integrate memory poisoning validation with existing tests. This has two main advantages:
* All of the tests are written already, potentially saving development time.
* The code coverage of these tests is already much greater than would be achievable writing new tests from scratch.
In an ideal world, we would be able to take a grand, encompassing approach whereby we would simply replace the implementation of `mbedtls_calloc()` and all tests would transparently run with memory poisoning enabled. Unfortunately, there are some significant difficulties with this idea:
* We cannot automatically distinguish which allocated buffers are shared buffers that need memory poisoning enabled.
* Some input buffers to tested functions may be stack allocated so cannot be poisoned automatically.
Instead, consider a more modest strategy. Create a function:
Further interface design or careful use of macros may make this a little less cumbersome than it seems in this example.
The poison copying functions should be written so as to evaluate to no-ops based on the value of a config option. They also need not be added to all tests, only to a 'covering set' of important tests.
##### Validation with new tests
Validation with newly created tests is initially simpler to implement than using the existing tests, since the tests can know about memory poisoning from the start. However, re-implementing testing for most PSA interfaces (even only basic positive testing) is a large undertaking. Furthermore, not much is gained over the previous approach, given that it seems straightforward to wrap PSA function calls in existing tests with poisoning code.
**Design decision: Add memory poisoning code to existing tests rather than creating new ones, since this is simpler and produces greater coverage.**
#### Discussion
Of all discussed approaches, validation by memory poisoning appears as the best. This is because it:
* Does not require complex linking against different versions of `malloc()` (as is the case with the memory pool approach).
* Allows automated testing (unlike the review approach).
**Design decision: Use a memory poisoning approach to validate copying.**
TODO: when there is a requirement on drivers, how to we validate that our built-in implementation meets these requirements? (This may be through testing, review, static analysis or any other means or a combination.)
Note: focusing on read-read inconsistencies for now, as most of the cases where we aren't copying are inputs.
#### Linux mprotect+ptrace
Idea: call `mmap` to allocate memory for arguments and `mprotect` to deny or reenable access. Use `ptrace` from a parent process to react to SIGSEGV from a denied access. On SIGSEGV happening in the faulting region:
1. Use `ptrace` to execute a `mprotect` system call in the child to enable access. TODO: How? `ptrace` can modify registers and memory in the child, which includes changing parameters of a syscall that's about to be executed, but not directly cause the child process to execute a syscall that it wasn't about to execute.
2. Use `ptrace` with `PTRACE_SINGLESTEP` to re-execute the failed load/store instrution.
3. Use `ptrace` to execute a `mprotect` system call in the child to disable access.
4. Use `PTRACE_CONT` to resume the child execution.
Record the addresses that are accessed. Mark the test as failed if the same address is read twice.
#### Debugger + mprotect
Idea: call `mmap` to allocate memory for arguments and `mprotect` to deny or reenable access. Use a debugger to handle SIGSEGV (Gdb: set signal catchpoint). If the segfault was due to accessing the protected region:
1. Execute `mprotect` to allow access.
2. Single-step the load/store instruction.
3. Execute `mprotect` to disable access.
4. Continue execution.
Record the addresses that are accessed. Mark the test as failed if the same address is read twice. This part might be hard to do in the gdb language, so we may want to just log the addresses and then use a separate program to analyze the logs, or do the gdb tasks from Python.
The best approach for validating the correctness of memory accesses is an open question that requires further investigation and prototyping. The above sections discuss some possibilities.
However, there is one additional consideration that may make this easier. The careful-access approach to memory protection is only planned for hash and MAC algorithms. These lend themselves to a linear access pattern on input data; it may be simpler to test that a linear pattern is followed, rather than a random-access single-access-per-location pattern.
TODO: analyze the built-in implementations of mechanisms for which there is a requirement on drivers. By code inspection, how satisfied are we that they meet the requirement?
For efficiency, we are likely to want mechanisms to bypass the copy and process buffers directly in builds that are not affected by shared memory considerations.
Expand this section to document any mechanisms that bypass the copy.
Make sure that such mechanisms preserve the guarantees when buffers overlap.
Asymmetric signature | Careful access | Copying | Inputs to signatures are passed to a hash. This will no longer hold once PureEdDSA support is implemented.
Asymmetric encryption | Copying | Copying |
Key agreement | Copying | Copying |
PAKE | Copying | Copying |
Key import / export | Copying | Copying | Keys may be imported and exported in DER format, which is a structured format and therefore susceptible to read-read inconsistencies and potentially write-read inconsistencies.
These seem to be a repeat of the same function, however it is useful to retain two separate functions for input and output parameters so that we can use different test hooks in each when using memory poisoning for tests.
Given that the majority of functions will be allocating memory on the heap to copy, it may help to build convenience functions that allocate the memory as well. One function allocates and copies the buffers:
This function would first copy the `buffers->out` buffer to the user-supplied output buffer and then free `buffers->inp` and `buffers->out`.
Some PSA functions may not use these convenience functions as they may have local optimizations that reduce memory usage. For example, ciphers may be able to use a single intermediate buffer for both input and output.
### Validation of copying
As discussed above, the best strategy for validation of copies appears to be validation by memory poisoning.
To implement this validation, we need several things:
1. The ability to allocate memory in individual pages.
2. The ability to poison memory pages in the copy functions.
3. Tests that exercise this functionality.
We can implement (1) as a test helper function that allocates full pages of memory so that we can safely set permissions on them:
This allocates a buffer of the requested size that is guaranteed to lie entirely within its own memory page. It also calls `mprotect()` so that the page is inaccessible.
This function should call `mprotect()` on the buffer to prevent it from being accessed (when `poisoned == 1`) or to allow it to be accessed (when `poisoned == 0`). Note that `mprotect()` requires a page-aligned address, so the function may have to do some preliminary work to find the correct page-aligned address that contains `buffer`.
