This patch does several things.
First, `SlidingSync::on_list` is now async, and accept async closures.
Second, `SlidingSync::lists` and `::rooms` are behind an `AsyncRwLock`
instead of a `StdRwLock`. The rest of the patch updates the consequence
of this.
This patch replaces a panic (`Option::expect`) by a `tracing::error`
log.
Imagine the Sliding Sync proxy responds with the following payload:
```json
{
"pos": …,
"lists": {
…: {
"count": …,
"ops": [
{
"op": "INSERT",
"index": 0,
"room_id": !foo:bar.org",
}
]
}
},
"rooms": []
}
```
It's an invalid response, as it will update the room list entry (because
it's present in `lists.$list.ops`), but the room won't be created (it's
absent in the `rooms`).
This situation creates a panic in the patched code. We don't want to
crash if the server replies with invalid data.
Ultimately, we should check that the sync operations are valid regarding
the rooms' updates.
This patch uses `FutureExt::now_or_never` so that we don't wait on the
future to be resolved to get a result. If we have a bug in our code and
the test expects a value, the test will hang forever, which is not a
desired behavior. With `now_or_never`, this situation cannot happen.
The only reason why a sender can fail to send a message is because there
is no receiver. In the current design of `SlidingSync`, there is only
one receiver active per sync-loop. Thus, calling `SlidingSync::add_list`
may fail if `sync` has not been started. Hence the need for a new
`internal_channel_send_if_possible` method which will never fail: it
will send a message is possible, otherwise it won't do anything.
This turns more functions infallible.
`tokio::sync::broadcast` is interesting as it's possible to
generate receivers per subscribers. Being able to do that allows
us to remove the new for an `AsyncLock` around the receiver in
`SlidingSync::internal_channel`, and it can even remove the need to hold
a receiver entirely!
Another improvement is that new receivers can't receive past messages,
so we no longer need to drain the internal channel.
Another improvement is that the sender' `send` method is synchronous!
Which helps to make many functions no longer `async`.
Something weird is happening with ElementX iOS. When
`SlidingSync::stop_sync` is called, the internal message `SyncLoopStop`
has time to be sent in the internal channel, but iOS decides to suspend
the code before the sync-loop processes it. When ElementX decides to
start the sync-loop again, it immediately processes the `SyncLoopStop`
message, and… stops the sync-loop.
This patch ensures that `SlidingSync::sync` drains the internal channel
before starting the sync-loop for real. `tokio::sync::mpsc::Receiver`
type has no `drain` method, so this patch implements its
own logic by calling `try_recv` in a loop, until it returns
`Err(TryRecvError::Empty)`.
* feat: introduce SlidingSyncSelectiveModeBuilder
* feat: get rid of `CannotModifyRanges` \o/
* chore: rustfmt
* chore: remove scope in test
* chore: move request generator update to its own mod, gets rid of `set_ranges`
* chore: add comments on the request generator methods
* chore: sink one range_end definition into the match arm that matters
* ffi: remove unused `add_range` method
* test: update incorrect test expectation
* chore: make clippy happy
* address first review comments
* review: don't reuse the ranges field for two things
* chore: use a builder pattern for sliding sync selective mode
* address review comments + CI
Signed-off-by: Benjamin Bouvier <public@benj.me>
