The new in-`ChannelManager` retries logic does retries as two
separate steps, under two separate locks - first it calculates
the amount that needs to be retried, then it actually sends it.
Because the first step doesn't udpate the amount, a second thread
may come along and calculate the same amount and end up retrying
duplicatively.
Because we generally shouldn't ever be processing retries at the
same time, the fix is trivial - simply take a lock at the top of
the retry loop and hold it until we're done.
In anticipation of the next commit(s) adding threaded tests, we
need to ensure our lockorder checks work fine with multiple
threads. Sadly, currently we have tests in the form
`assert!(mutex.try_lock().is_ok())` to assert that a given mutex is
not locked by the caller to a function.
The fix is rather simple given we already track mutexes locked by a
thread in our `debug_sync` logic - simply replace the check with a
new extension trait which (for test builds) checks the locked state
by only looking at what was locked by the current thread.
We're no longer supporting manual retries since
ChannelManager::send_payment_with_retry can be parameterized by a retry
strategy
This commit also updates all docs related to retry_payment and abandon_payment.
Since these docs frequently overlap with changes in preceding commits where we
start abandoning payments on behalf of the user, all the docs are updated in
one go.
Documentation for CustomMessageHandler wasn't clear how it is related to
PeerManager and contained some grammatical and factual errors. Re-write
the docs and link to the lightning_custom_message crate.
When a peer disconnects but still has channels, the peer's `peer_state`
entry in the `per_peer_state` is not removed by the `peer_disconnected`
function. If the channels with that peer is later closed while still
being disconnected (i.e. force closed), we therefore need to remove the
peer from `peer_state` separately.
To remove the peers separately, we push such peers to a separate HashSet
that holds peers awaiting removal, and remove the peers on a timer to
limit the negative effects on parallelism as much as possible.
Updates multiple instances of the `ChannelManager` docs related to the
previous change that moved the storage of the channels to the
`per_peer_state`. This docs update corrects some grammar errors and
incorrect information, as well as clarifies documentation that was
confusing.
When we apply the new gossip-async-check backpressure on peer
connections, if a peer has never sent us a `channel_announcement`
at all, we really shouldn't delay reading their messages.
This does so by tracking, on a per-peer basis, whether they've sent
us a channel_announcement, and resetting that state whenever we're
not backlogged.
Now that the `RoutingMessageHandler` can signal that it needs to
apply message backpressure, we implement it here in the
`PeerManager`. There's not much complicated here, aside from noting
that we need to add the ability to call `send_data` with no data
to indicate that reading should resume (and track when we may need
to make such calls when updating the routing-backpressure state).
Now that we allow `handle_channel_announcement` to (indirectly)
spawn async tasks which will complete later, we have to ensure it
can apply backpressure all the way up to the TCP socket to ensure
we don't end up with too many buffers allocated for UTXO
validation.
We do this by adding a new method to `RoutingMessageHandler` which
allows it to signal if there are "many" checks pending and
`channel_announcement` messages should be delayed. The actual
`PeerManager` implementation thereof is done in the next commit.
Gossip messages which were verified against the chain
asynchronously should still be forwarded to peers, but must now go
out via a new `P2PGossipSync` parameter in the
`AccessResolver::resolve` method, allowing us to wire them up to
the `P2PGossipSync`'s `MessageSendEventsProvider` implementation.
When we process gossip messages asynchronously we may find that we
want to forward a gossip message to a peer after we've returned
from the existing `handle_*` method. In order to do so, we need to
be able to send arbitrary loose gossip messages back to the
`PeerManager` via `MessageSendEvent`.
This commit modifies `MessageSendEvent` in order to support this.
If we have a `channel_announcement` which is waiting on a UTXO
lookup before we can process it, and we receive a `channel_update`
or `node_announcement` for the same channel or a node which is a
part of the channel, we have to wait until the lookup completes
until we can decide if we want to accept the new message.
Here, we store the new message in the pending lookup state and
process it asynchronously like the original `channel_announcement`.
If we receive two `channel_announcement`s for the same channel at
the same time, we shouldn't spawn a second UTXO lookup for an
identical message. This likely isn't too rare - if we start syncing
the graph from two peers at the same time, it isn't unlikely that
we'll end up with the same messages around the same time.
In order to avoid this we keep a hash map of all the pending
`channel_announcement` messages by SCID and simply ignore duplicate
message lookups.
