Making build times considerably faster.
On my machine, after running `cargo clean`, `cargo build -p swap`
takes 2min 19s.
The updated dependency also comes with a critical fix to the `Scalar`
type, which originally wrongly assumed that secp256k1 and ed25519
scalars had the same endianness. For this reason, we now have to
reverse the bytes of recovered scalars if we are to use them on a
different chain.
Finally, there is no need to append `RUST_MIN_STACK=100000000` to
avoid stack overflows in tests and when running the binary.
Unfortunately, I had to put the wrap the swarm in Alice's `Network`
struct in an `Arc<Mutex<T>>` in order to be able to use `backoff` to
control the retry mechanism. This is because the stream of events
cannot be turned into a `SharedFuture` (unlike Bob's).
It would be good to find an alternative solution.
Also:
- Move generator functions to `alice` and `bob` modules. This makes
using `tracing` a lot easier, since the context of the file name let's
us differentiate between Alice's and Bob's generator logs more
clearly.
- Accept 0 confirmations when watching for the Monero lock
transaction. This should eventually be configured by the application,
but in the tests it's making things unexpectedly slower.
- Make it the same for Alice and Bob.
- Make it contain a wallet client instead of the `Monero` struct.
Also:
Remove `Container` from inside `Monero` struct. The caller of `new`
can simply ensure that `Container` is not dropped to keep the
container alive.
This makes the `Monero` struct easier to work with, as it just holds
the data necessary to create the different clients created during
`init`, and does not have any lifetime restrictions.
By having two modules that declare and use the `harness` module we ran
into a situation where not all parts of the `harness` module were
being used by both of these other test modules. This was producing
annoying warnings.
For now this can be solved by marking the `harness` module declaration
as public. This has no real effect since modules under `/tests` cannot
be depended on elsewhere, but it satisfies the compiler. It does mean
that we will not hear about unused functions, but this change should
be temporary and it's a very minor problem given that it's just test
code.
Mimics what @thomaseizinger did here [1] and here [2].
This has the advantage that the consumer has more freedom to execute
`Action`s without having to implement particular traits. The error
handling required inside this protocol-executing function is also
reduced.
As discussed with Thomas, for this approach to work well, the
trait functions such as `receive_transfer_proof` should be infallible,
and the implementer should be forced to hide IO errors behind a retry
mechanism.
All of these asynchronous calls need to be "raced" against
the abort condition (determined by the `refund_timelock`), which is
missing in the current state of the implementation.
The initial handshake of the protocol has not been included here,
because it may not be easy to integrate this approach with libp2p, but
a couple of messages still need to exchanged. I need @tcharding to
tell me if it's feasible/good to do it like this.
[1]
https://github.com/comit-network/comit-rs/blob/move-nectar-swap-to-comit/nectar/src/swap/comit/herc20_hbit.rs#L57-L184.
[2] e584d2b14f/nectar/src/swap.rs (L716-L751).
