Vodozemac used to accept and return strings when encrypting and
decrypting. This is quite unusual for a pure cryptographic library so we
switched towards the usual setup where we encrypt/decrypt raw bytes.
Since we do encrypt/decrypt JSON strings in Matrix land, we do the
string conversions over here.
In async functions, the Node.js GC may or may not (that's a random
behavior) collect the arguments passed to the function as soon as it
returns. The function may not be executed yet, since it's async. Thus,
it leads to memory corruption: The function tries to read later on the
value inside an argument and… it crashes at best.
To avoid this bug, there is no other choice than cloning the values
before the function returns, in its “sync path” (so before any
transformation of an `.await` point into an “async block”).
The performance impact is not “massive”, I'm not sure it could be
noticeable easily since it is most of the time related to identifiers
(e.g. `UserId`), which are cheap to clone. I have to find the balance
here, and cloning offers the best trade off from my point of view.
So, Clippy suggests to change `(&member).into()` to `member.into()`
but it's the same, and `From<T>` is not implemented for this `T`, only
`From<&T>` is present. Thus, to deceive Clippy, I'm using
`std::borrow::Borrow` here. Not super happy with that though…
Clippy on nigtly is raising a warning, which is turned into an error
on the CI. It's the [`explicit_auto_deref`
lint](https://rust-lang.github.io/rust-clippy/master/index.html#explicit_auto_deref). I
suspect it's a false-positive but I'm not sure. Anyway, to workaround
this and unblock our CI, let's call `DerefMut::deref_mut` manually:
it's clearer anyway.
This patch first implements the new `Signatures`, `Signature` and `MaybeSignature` types.
Then, it moves some Vodozemac types into their own module, and
implements the new `Ed25519Signature` type.
Finally, it implements `OlmMachine.sign`.
For some unknown reasons, sometimes, randomly, one test (initializing
an `OlmMachine` with a local store with a passphrase) can take more
than 5s, only on Github Actions. Let's increase the test timeout value
so that the entire test suite doesn't fail.
Charles Wright