On 8 December 2017 at 23:11, Eric Biggers [off-list ref] wrote:

On Fri, Dec 08, 2017 at 10:54:24PM +0000, Ard Biesheuvel wrote:

quoted

quoted

quoted

Note that there are two conflicting conventions for what inputs ChaCha takes.
The original paper by Daniel Bernstein
(https://cr.yp.to/chacha/chacha-20080128.pdf) says that the block counter is
64-bit and the nonce is 64-bit, thereby expanding the key into 2^64 randomly
accessible streams, each containing 2^64 randomly accessible 64-byte blocks.

The RFC 7539 convention is equivalent to seeking to a large offset (determined
by the first 32 bits of the 96-bit nonce) in the keystream defined by the djb
convention, but only if the 32-bit portion of the block counter never overflows.

Maybe it is only RFC 7539 that matters because that is what is being
standardized by the IETF; I don't know.  But it confused me.

The distinction only matters if you start the counter at zero (or
one), because you 'lose' 32 bits of IV that will never be != 0 in
practice if you use a 64-bit counter.

So that argues for not exposing the block counter as part of the API,
given that it should start at zero anyway, and that you should take
care not to put colliding values in it.

quoted

Anyway, I actually thought it was intentional that the ChaCha implementations in
the Linux kernel allowed specifying the block counter, and therefore allowed
seeking to any point in the keystream, exposing the full functionality of the
cipher.  It's true that it's easily misused though, so there may nevertheless be
value in providing a nonce-only variant.

Currently, the skcipher API does not allow such random access, so
while I can see how that could be a useful feature, we can't really
make use of it today. But more importantly, it still does not mean the
block counter should be exposed to the /users/ of the skcipher API
which typically encrypt/decrypt blocks that are much larger than 64
bytes.

... but now that I think of it, how is this any different from, say,
AES in CTR mode? The counter is big endian, but apart from that, using
IVs derived from a counter will result in the exact same issue, only
with a shift of 16 bytes.

That means using file block numbers as IV is simply inappropriate, and
you should encrypt them first like is done for AES-CBC

The problem with using a stream cipher --- whether that is ChaCha20, AES-CTR, or
something else --- for disk/file encryption is that by necessity of file/disk
encryption, each time the "same" block is written to, the IV is the same, which
is really bad for stream ciphers (but not as bad for AES-XTS, AES-CBC, etc.).
It's irrelevant whether you do ESSIV or otherwise encrypt the IVs.  ESSIV does
make the IV for each offset unpredictable by an attacker, which is desirable for
AES-CBC, but it doesn't stop the IV from being repeated for each overwrite.

I'm not suggesting using an encrypted IV to fix the stream cipher
issue, I'm well aware that that is impossible. What I am saying is
that the counter collision can be mitigated by encrypting the IV.

And just to clarify, you definitely *can* seek to any position in the ChaCha20
stream using the existing ChaCha20 implementations and the existing skcipher
API, simply by providing the appropriate IV.  Maybe it was unintentional, but it
does work.  chacha20poly1305.c even uses it to start at block 1 instead of block
0.  I don't know whether there are other users, though.

Well, I understand that that's how ChaCha20 works, and that you can
manipulate the IV directly to start at another point in the keystream.
AES-CTR can do exactly the same, for the same reason. What I am saying
is that the skcipher API does not allow you to decrypt an arbitrary
part of a block, which could benefit from the ability of not having to
generate the entire key stream.

So the more we discuss this, the more I think there is actually no
difference with AES-CTR (apart from the block size), and there a
similar enhancement in RFC3686 where the IV does not cover the AES
block level counter, making it safe to use another counter to generate
the IVs.

Of course, this is essentially what you did for the fscrypt code, I
just don't like seeing that kind of reasoning being implement in the
crypto API client.