bcache/bcachefs encryption design:
This document is intended for review by cryptographers and other experience implementers of cryptography code, before the design is frozen. Everything described in this document has been implemented and tested.
The code may be found at https://evilpiepirate.org/git/linux-bcache.git/log/?h=bcache-encryption
Feedback and comments should be sent to Kent Overstreet, firstname.lastname@example.org.
Main bcachefs page: Bcachefs
Bcachefs provides whole-filesystem encryption, using ChaCha20/Poly1305. Encryption may be enabled when creating a filesystem, or encryption may be enabled on an existing filesystem (TODO: implement interface for enabling encryption on an existing filesystem - kernel code exists).
$ bcache format --encrypted /dev/sda
(Enter passphrase when prompted)
$ bcache unlock /dev/sda
(Enter passphrase again)
Then mount as normal:
$ mount /dev/sda /mnt
Bcachefs encryption is meant to be a clean slate design that prioritizes security and robustness, and is meant to defend against a wider variety of adversarial models than is typical in existing filesystem level or block level encryption.
In particular, the goal is to be secure even when the attacker controls the storage device itself, and can see reads and writes as they happen and return arbitrary data from read requests.
Filesystem vs. directory encryption
We do not currently offer per directory encryption; instead, we take an "encrypt everything" approach.
With per directory encryption, preventing metadata from leaking is highly problematic. For example, file sizes - file sizes are required for fsck, so they would have to be stored unencrypted - or failing that some complicated way of deferring fsck for that part of the filesystem until the key has been provided. There would be additional complications around filenames, xattrs, extents (and inline extents), etc. - not necessarily insurmountable, but they would definitely lead to a more complicated, more fragile design.
With whole filesystem encryption, it’s much easier to say what is and isn’t encrypted, we can encrypt at the granularity of an entire metadata write (a journal entry, a btree node) and it's much easier to show that the contents - everything after the header for that particular metadata write - will not leak.
By virtue of working within a copy on write filesystem with provisions for ZFS style checksums (that is, checksums with the pointers, not the data), we’re able to use a modern AEAD style construction. We use ChaCha20 and Poly1305. We use the cyphers directly instead of using the kernel AEAD library. However, we do follow pretty closely the approach of RFC 7539.
Note that ChaCha20 is a stream cypher. This means that it’s critical that we use a cryptographic MAC (which would be highly desirable anyways), and also avoiding nonce reuse is critical. Getting nonces right is where most of the trickiness is involved in bcachefs’s encryption.
The current algorithm choices are not hard coded. Bcachefs already has selectable checksum types, and every individual data and metadata write has a field that describes the checksum algorithm that was used. On disk, encrypted data is represented as a new checksum type - so we now have [none, crc32c, crc64, chacha20/poly1305] as possible methods for data to be checksummed/encrypted. If in the future we add new encryption algorithms, users will be able to switch to the new algorithm on existing encrypted filesystems; new data will be written with the new algorithm and old data will be read with the old algorithm until it is rewritten.
Key derivation, master key
Userspace tooling takes the user's passphrase and derives an encryption key with scrypt. This key is made available to the kernel (via the Linux kernel's keyring service) prior to mounting the filesystem.
On filesystem mount, the userspace provided key is used to decrypt the master key, which is stored in the superblock - also with ChaCha20. The master key is encrypted with an 8 byte header, so that we can tell if the correct key was supplied.
TODO: Add a field to the superblock specifying the key derivation function, so that we can transition to newer KDFs later (e.g. Argon2) or specify cost parameters.
Except for the superblock, no metadata in bcache/bcachefs is updated in place - everything is more or less log structured. Only the superblock is stored unencrypted; other metadata is stored with an unencrypted header and encrypted contents.
The superblock contains:
- Label and UUIDs identifying the filesystem
- A list of component devices (for multi-device filesystems), and information on their size, geometry, status (active/failed), last used timestamp
- Filesystem options
- The location of the journal
For the rest of the metadata, the unencrypted portion contains:
- 128 bit checksum/MAC field
- Magic number - identifies a given structure as btree/journal/allocation information, for that filesystem
- Version number (of on disk format), flags (including checksum/encryption type).
- Sequence numbers: journal entries have an ascending 64 bit sequence number, btree node entries have a random 64 bit sequence number identifying them as belonging to that node. Btree nodes also have a field containing the sequence number of the most recent journal entry they contain updates from; this is stored unencrypted so it can be used as part of the nonce.
- Size of the btree node entry/journal entry, in u64s
Btree node layout information is encrypted; an attacker could tell that a given location on disk was a btree node, but the part of the header that indicates what range of the keyspace, or which btree ID (extents/dirents/xattrs/etc.), or which level of the btree is all encrypted.
Journal entries use their sequence number - which is unique for a given filesystem. When metadata is being replicated and we're doing multiple journal writes with the same sequence number - and thus nonce - we really are writing the same data (we only checksum once, not once per write).
Btree nodes concatenate a few things for the nonce:
- A 64 bit random integer, which is generated per btree node (but btree nodes are log structured, so entries within a given btree node share the same integer).
- A journal sequence number. For btree node writes done at around the same point in time, this field can be identical in unrelated btree node writes - but only for btree nodes writes done relatively close in time, so the journal sequence number plus the previous random integer should be more than sufficient entropy.
- And lastly the offset within the btree node, so that btree node entries sharing the same random integer are guaranteed a different nonce.
Allocation information (struct prio_set): bcache/bcachefs doesn't have allocation information persisted like other filesystems, but this is our closest equivalent - this structure mainly stores generation numbers that correspond to extent pointers.
Allocation information uses a dedicated randomly generated 96 bit nonce field.
Data writes have no unencrypted header: checksums/MACs, nonces, etc. are stored with the pointers, ZFS style.
Bcache/bcachefs is extent based, not block based: pointers point to variable sized chunks of data, and we store one checksum/MAC per extent, not per block: a checksum or MAC might cover up to 64k (extents that aren't checksummed or compressed may be larger). Nonces are thus also per extent, not per block.
By default, for data extents the Poly1305 MAC is truncated to 80 bits, for space efficiency reasons. Optionally the full 128 bit macs may be stored, at the cost increasing the size of extents by 8 bytes (with 80 bit macs, an extent with a single replica will typically be 32 bytes, or 40 bytes with 128 bit macs).
This should be completely safe for the vast majority of uses cases. Most uses of cryptographic MACs are in networked applications, where an attacker may be able to send an unlimited number of forged messages: in that environment, a 64 bit mac is clearly insufficient - if an attacker is able to send 232 forgery attempts (not a huge number these days), probability of success is 1 / 232 - which is not considered a remotely safe margin by cryptographers.
However, with a filesystem, even in the case of a completely compromised device (say an attacker has compromised the firmware on the disk, and is able to return whatever they want when we read a sector) - if the MAC doesn't match (because the attacker is attempting to forge data), we consider the device to be failing and very shortly we're going to stop using it - we won't attempt to reread data that appears to be corrupt indefinitely. So, attacker gets a very small (on the order of 10) attempts to forge a particular extent. In the very worst case, if we're trying very hard to migrate data off a device that appears to be bad, the attacker might get ~10 attempts multiplied by the number of extents on the device - but the number of forgery attempts should be clearly bounded.
If the user is in an environment where transient failures/corruption are expected should be tolerated, instead of assuming the device is bad (e.g. the disks are accessed over the network, and the network path is known to corrupt data) - in that situation 128 bit macs should be used (and in the future we may enforce that if the maximum number of read retries is set to more than a small number, 128 bit macs must be used if encryption is in use).
We don't wish to simply add a random 96 bit nonce to every extent - that would inflate our metadata size by a very significant amount. Instead, keys (of which extents are a subset) have a 96 bit version number field; when encryption is enabled, we ensure that version numbers are enabled and every new extent gets a new, unique version number.
However, extents may be partially overwritten or split, and then to defragment we may have to rewrite those partially overwritten extents elsewhere. We cannot simply pick a new version number when we rewrite an extent - that would break semantics other uses of version numbers expect.
When we rewrite an extent, we only write the currently live portions of the extent - we don't rewrite the parts that were overwritten. We can't write it out with the same nonce as the original extent.
If we concatenated the version number with the offset within the file, and the extent's current size - that would work, except that it would break fcollapse(), which moves extents to a new position within a file. We are forced to add some additional state to extents.
We could add a small counter that is incremented every time the size of an extent is reduced (and the data it points to changes); we can easily bound the size of the counter we need by the maximum size of a checksummed extent. But this approach fails when extents are split.
What can work is if we add a field for "offset from the start of the original extent to the start of the current extent" - updating that field whenever we trim the front of an extent.
If we have that, then we could simply skip ahead in the keystream to where the currently live data lived in the original extent - there's no problem with nonce reuse if you're encrypting exactly the same data. Except - that fails with compression, since if we take an extent, drop the first 4k, and compress it, that won't give the same data as if we compress it and then drop the first 4k of the compressed data.
The approach almost works though, if we take that offset and use it as part of our nonce: what we want to do is construct a function that will output the same nonce iff two extents (fragments of the same original extent) really are the same data.
Offset into the original extent works in the absence of compression - two fragments with the same offset but different sizes will be equal in their common prefix, ignoring compression. We can handle compression if we also include both the current size, and the current compression function - offset + current size uniquely determines the uncompressed data, so, offset + current size + compression function will uniquely determine the compressed output.
Nonce reuse on startup
After recovery, we must ensure we don't reuse existing version numbers - we must ensure that newly allocated version numbers are strictly greater than any version number that has every been used before.
The problem here is that we use the version number to write the data before adding the extent with that version number to the btree: after unclean shutdown, there will have been version numbers used to write data for which we have no record in the btree.
The rigorous solution to this is to add a field (likely to the journal header) that indicates version numbers smaller than that field may have been used. However, we don't do that yet - it's not completely trivial since it'll add another potential dependency in the IO path that needs some analysis.
The current solution implemented by the code is to scan every existing version number (as part of an existing pass), and set the next version number to allocate to be 64k greater than the highest existing version number that was found.