zxcrypt is a block device filter driver that transparently encrypts data being written to and decrypts being read from data a block device. The underlying block device that a zxcrypt device uses may be almost any block device, including raw disks, ramdisks, GPT partitions, FVM partitions or even other zxcrypt devices. The only restriction is that the block size be page-aligned. Once bound, the zxcrypt device will publish another block device in the device tree that consumers can interact with normally.
zxcrypt contains both a driver and library
Provided by libzxcrypt.so are four functions for managing zxcrypt devices. Each takes one or more
zxcrypt_key_t keys, which associates the key data, length, and slot in the case of multiple keys.
- The zxcrypt_format function takes an open block device, and writes the necessary encrypted metadata to make it a zxcrypt device. The zxcrypt key provided does not protect the data on the device directly, but is used to protect the data key material.
zx_status_t zxcrypt_format(int fd, const zxcrypt_key_t* key);
- The zxcrypt_bind function instructs the driver to read the encrypted metadata and extract the data key material to use in transparently transforming I/O data.
zx_status_t zxcrypt_bind(int fd, const zxcrypt_key_t *key);
- The zxcrypt_rekey function uses the old key to first read the encrypted metadata, and the new key to write it back.
zx_status_t zxcrypt_rekey(int fd, const zxcrypt_key_t* old_key, const zxcrypt_key_t* new_key);
- The zxcrypt_shred function first verifies that the caller can access the data by using the key provided to read the encrypted metadata. If this succeeded, it then destroys the encrypted metadata containing the data key material. This prevents any future access to the data.
zx_status_t zxcrypt_shred(int fd, const zxcrypt_key_t* key);
zxcrypt is written as a DDKTL device driver. src/lib/ddktl is a C++ framework for writing drivers in Fuchsia. It allows authors to automatically supply the src/lib/ddk function pointers and callbacks by using templatized mix-ins.
There are two small pieces of functionality which cannot be written in DDKTL and C++:
- The driver binding logic, written using the C preprocessor macros of DDK's binding.h.
- The completion routines of ulib/sync, which are used for synchronous I/O and are incompatible with C++ atomics.
The device starts worker threads that run for the duration of the device and create a pipeline for all I/O requests. Each has a type of I/O it operates on, a queue of incoming requests I/O that it will wait on, and a data cipher. When a request is received, if the opcode matches the one it is looking for, it will use its cipher to transform the data in the request before passing it along.
The overall pipeline is as shown:
DdkIotxnQueue -+ \ Worker 1: Underlying Worker 2: Original BlockRead ---+---> Encrypter ---> Block ---> Decrypter ---> Completion / Acts on writes Device Acts on reads Callback BlockWrite -+
The "encrypter" worker encrypts the data in every I/O write request before sending it to the underlying block device, and the "decrypter" worker decrypts the data in every I/O read response coming from the underlying block device. The cipher must have a key length of at least 16 bytes, be semantically secure (IND-CCA2) and incorporate the block offset as a "tweak". Currently, AES256-XTS is in use.
Rings and Txns
In order to keep the encryption and decryption of data transparent to original I/O requester, the workers must copy the data when transforming it. The I/O request sent through the pipeline is not actually the original request, but instead a "shadow" request that encapsulates the original request.
As shadow requests are needed, they are allocated backed sequentially by pages in the VMO. When the worker needs to transform the data it either encrypts data from the original, encapsulated write request into the shadow request, or decrypts data from the shadow request into the original, encapsulated read request. As soon as the original request can be handed back to the original requester, the shadow request is deallocated and its page decommitted. This ensures no more memory is used than is needed for outstanding I/O requests.
The key material for encrypting and decrypting the data is referred to as the data key, and is
stored in a reserved portion of the device called the
superblock. The presence of this superblock
is critical; without it, it is impossible to recreate the data key and recover the data on the
device. As a result, the superblock is copied to multiple locations on the device for redundancy.
These locations are not visible to zxcrypt block device consumers. Whenever the zxcrypt driver
successfully reads and validates a superblock from one location, it will copy this to all other
superblock locations to help "self-heal" any corrupted superblock locations.
The superblock format is as follows, with each field described in turn:
+----------------+----------------+----+-----...-----+----...----+------...------+ | Type GUID | Instance GUID |Vers| Sealed Key | Reserved | HMAC | | 16 bytes | 16 bytes | 4B | Key size | ... | Digest length | +----------------+----------------+----+-----...-----+----...----+------...------+
- Type GUID: Identifies this as a zxcrypt device. Compatible with GPT.
- Instance GUID: Per-device identifier, used as the KDF salt as explained below.
- Version: Used to indicate which cryptographic algorithms to use.
- Sealed Key: The data key, encrypted by the wrap key as described below.
- Reserved: Unused data to align the superblock with the block boundary.
- HMAC: A keyed digest of the superblock up to this point (including the Reserved field).
The wrap key, wrap IV, and HMAC key are all derived from a KDF. This KDF is an RFC 5869 HKDF, which combines the key provided, the "salt" of the instance GUID and a per-use label such as "wrap" or "hmac". The KDF does NOT try to do any rate-limiting. The KDF mitigates the risk of key reuse, as a new random instance salt will lead to new derived keys. The HMAC prevents accidental or malicious modification to go undetected, without leaking any useful information about the zxcrypt key.
_NOTE: The KDF does NOT do any key stretching. It is assumed that an attacker can remove a device and attempt the key derivations on their own, bypassing the HMAC check and any possible rate limits. To prevent this, zxcrypt consumers should include properly rate-limited device keys, e.g. those from a TPM, in deriving their zxcrypt key._
There are a number of areas where further work could, should, or must be done:
Surface hidden bind failures
zxcrypt_bindmay indicate success even though the device fails to initialize. zxcrypt is NOT synchronously adding the device to the device tree when the binding logic is run. It must do I/O and cannot block the call to
device_bindfrom returning, so it spawn an initializer thread and adds the device when complete.
As of 10/2017, this is an active area of DDK development and the policy is changing to requiring the device to be added before return, with an additional call to publish that may come later. With this it may be desirable to have the call to
zxcrypt_bindblock synchronously for callers until the device is ready or has unambiguously failed to bind.
Use AEAD instead of AES-XTS
It is widely recognized that AEADs provide superior cryptographic protection by validating the integrity of their data before decrypting it. This is desirable, but requires additional per-block overhead. This means either that consumers will need to consume non-page-aligned blocks (once the in-line overhead is removed), or zxcrypt will need to store the overhead out-of-line and handle non-atomic write failures.
Support multiple keys
To facilitate key escrow and/or recovery, it is straightforward to modify the superblock format to have a series of cryptographic envelopes. In anticipation of this, the libzxcrypt API takes a variable number of keys, although the only length currently supported is 1, and the only valid slot is 0.
Adjust number of workers
Currently there is one encrypter and one decrypter. These are designed to work with an arbitrary number of threads, so performance tuning may be need to find the optimal number of workers that balances I/O bandwidth with scheduler churn.
Remove internal checks
Currently, the zxcrypt code checks for many errors conditions at internal boundaries and returns informative errors if those conditions aren't met. For performance, those that arise from programmer error only and not data from either the requester or underlying device could be converted to "debug" assertions that are skipped in release mode.