This fixes a bug introduced when we added heartbeat support. Recall that we
handle the Bitcoin connection state machine on a per-peer basis. Each
connection has a task created from the `Connection` struct, and a `Client:
tower::Service` "frontend" that passes requests to it via a channel. In the
`Connection` event loop, the connection checks whether the request channel has
been closed, indicating no further requests from the `Client`, in which case it
shuts itself down and cleans up resources. This occurs when all of the senders
have been dropped.
However, this behavior broke when we introduced heartbeat support, because we
spawned an additional task to send heartbeat messages along the request
channel. This meant that instead of having a single sender, dropped by the
`Client`, we have two senders, the `Client` and the "shadow client" task that
generates heartbeat messages. This means that when the `Client` is dropped, we
still have a live sender and the connection is not closed. To fix this, the
`Client` now uses a `oneshot` to shut down its corresponding heartbeat task.
This closes all senders.
Each subsection has to have `serde(default)` to get the behaviour we want
(delete all fields except the ones that have been changed); otherwise, we can
delete only entire sections.
Prior to this change, we required that services that are canceled do not
have a cancel handle in the `cancel_handles` list, based on the
assumption that the handle must have been removed in the process of
canceling this service.
This doesn't holding up though, because it is currently possible for us
to have the same peer connect to us multiple times, the second connect
removes the cancel handle of the original connect and inserts it's own
cancel handle in its place. In this scenario, when the first service is
polled for readiness it will see that it has been canceled and go to
clean itself up, but when it asserts that it doesn't have a cancel
handle it will see the cancel handle of the second connect event, which
uses the same key as the first connect, and fail its debug assertion.
This change removes that debug assert on the assumption that it is okay
for a peer to connect multiple times consecutively, and that the correct
behavior in that case is to just cancel the first connection and
continue as normal.
Prior to this change, the service returned by `zebra_network::init` would spawn background tasks that could silently fail, causing unexpected errors in the zebra_network service.
This change modifies the `PeerSet` that backs `zebra_network::init` to store all of the `JoinHandle`s for each background task it depends on. The `PeerSet` then checks this set of futures to see if any of them have exited with an error or a panic, and if they have it returns the error as part of `poll_ready`.
Co-authored-by: Jane Lusby <jane@zfnd.org>
Prior to this change, the seed subcommand would consistently encounter a panic in one of the background tasks, but would continue running after the panic. This is indicative of two bugs.
First, zebrad was not configured to treat panics as non recoverable and instead defaulted to the tokio defaults, which are to catch panics in tasks and return them via the join handle if available, or to print them if the join handle has been discarded. This is likely a poor fit for zebrad as an application, we do not need to maximize uptime or minimize the extent of an outage should one of our tasks / services start encountering panics. Ignoring a panic increases our risk of observing invalid state, causing all sorts of wild and bad bugs. To deal with this we've switched the default panic behavior from `unwind` to `abort`. This makes panics fail immediately and take down the entire application, regardless of where they occur, which is consistent with our treatment of misbehaving connections.
The second bug is the panic itself. This was triggered by a duplicate entry in the initial_peers set. To fix this we've switched the storage for the peers from a `Vec` to a `HashSet`, which has similar properties but guarantees uniqueness of its keys.
- Add a total peers metric to prevent races between measurements of
ready/unready peers (which can cause the sum to be wrong).
- Add an outbound request counter.
tower-buffer uses tokio's mpsc channels, not the futures-rs mpsc channels.
Unlike futures-rs mpsc channels, which have capacity n+m, where n is the buffer
size and m is the number of senders, tokio channels always have buffer size n.
This means that the buffer size is shared across all peer set handles.
Thanks to @hawkw for sharing details of the Tokio internals!
Previously, we relied on the owner of the handshake future to drive it to
completion. This meant that there were cases where handshakes might never be
completed, just because nothing was actively polling them.
The previous outbound peer connection logic got requests to connect to new
peers and processed them one at a time, making single connection attempts
and retrying if the connection attempt failed. This was quite slow, because
many connections fail, and we have to wait for timeouts. Instead, this logic
connects to new peers concurrently (up to 50 at a time).
Bitcoin does this either with `getblocks` (returns up to 500 following block
hashes) or `getheaders` (returns up to 2000 following block headers, not
just hashes). However, Bitcoin headers are much smaller than Zcash
headers, which contain a giant Equihash solution block, and many Zcash
blocks don't have many transactions in them, so the block header is
often similarly sized to the block itself. Because we're
aiming to have a highly parallel network layer, it seems better to use
`getblocks` to implement `FindBlocks` (which is necessarily sequential)
and parallelize the processing of the block downloads.
This doesn't clean the warnings about unused items in the builder, since
those are unused for a reason (the implementation that should use them
is missing).