[litep2p](https://github.com/altonen/litep2p) is a libp2p-compatible P2P
networking library. It supports all of the features of `rust-libp2p`
that are currently being utilized by Polkadot SDK.
Compared to `rust-libp2p`, `litep2p` has a quite different architecture
which is why the new `litep2p` network backend is only able to use a
little of the existing code in `sc-network`. The design has been mainly
influenced by how we'd wish to structure our networking-related code in
Polkadot SDK: independent higher-levels protocols directly communicating
with the network over links that support bidirectional backpressure. A
good example would be `NotificationHandle`/`RequestResponseHandle`
abstractions which allow, e.g., `SyncingEngine` to directly communicate
with peers to announce/request blocks.
I've tried running `polkadot --network-backend litep2p` with a few
different peer configurations and there is a noticeable reduction in
networking CPU usage. For high load (`--out-peers 200`), networking CPU
usage goes down from ~110% to ~30% (80 pp) and for normal load
(`--out-peers 40`), the usage goes down from ~55% to ~18% (37 pp).
These should not be taken as final numbers because:
a) there are still some low-hanging optimization fruits, such as
enabling [receive window
auto-tuning](https://github.com/libp2p/rust-yamux/pull/176), integrating
`Peerset` more closely with `litep2p` or improving memory usage of the
WebSocket transport
b) fixing bugs/instabilities that incorrectly cause `litep2p` to do less
work will increase the networking CPU usage
c) verification in a more diverse set of tests/conditions is needed
Nevertheless, these numbers should give an early estimate for CPU usage
of the new networking backend.
This PR consists of three separate changes:
* introduce a generic `PeerId` (wrapper around `Multihash`) so that we
don't have use `NetworkService::PeerId` in every part of the code that
uses a `PeerId`
* introduce `NetworkBackend` trait, implement it for the libp2p network
stack and make Polkadot SDK generic over `NetworkBackend`
* implement `NetworkBackend` for litep2p
The new library should be considered experimental which is why
`rust-libp2p` will remain as the default option for the time being. This
PR currently depends on the master branch of `litep2p` but I'll cut a
new release for the library once all review comments have been
addresses.
---------
Signed-off-by: Alexandru Vasile <alexandru.vasile@parity.io>
Co-authored-by: Dmitry Markin <dmitry@markin.tech>
Co-authored-by: Alexandru Vasile <60601340+lexnv@users.noreply.github.com>
Co-authored-by: Alexandru Vasile <alexandru.vasile@parity.io>
This commit introduces a new concept called `NotificationService` which
allows Polkadot protocols to communicate with the underlying
notification protocol implementation directly, without routing events
through `NetworkWorker`. This implies that each protocol has its own
service which it uses to communicate with remote peers and that each
`NotificationService` is unique with respect to the underlying
notification protocol, meaning `NotificationService` for the transaction
protocol can only be used to send and receive transaction-related
notifications.
The `NotificationService` concept introduces two additional benefits:
* allow protocols to start using custom handshakes
* allow protocols to accept/reject inbound peers
Previously the validation of inbound connections was solely the
responsibility of `ProtocolController`. This caused issues with light
peers and `SyncingEngine` as `ProtocolController` would accept more
peers than `SyncingEngine` could accept which caused peers to have
differing views of their own states. `SyncingEngine` would reject excess
peers but these rejections were not properly communicated to those peers
causing them to assume that they were accepted.
With `NotificationService`, the local handshake is not sent to remote
peer if peer is rejected which allows it to detect that it was rejected.
This commit also deprecates the use of `NetworkEventStream` for all
notification-related events and going forward only DHT events are
provided through `NetworkEventStream`. If protocols wish to follow each
other's events, they must introduce additional abtractions, as is done
for GRANDPA and transactions protocols by following the syncing protocol
through `SyncEventStream`.
Fixes https://github.com/paritytech/polkadot-sdk/issues/512
Fixes https://github.com/paritytech/polkadot-sdk/issues/514
Fixes https://github.com/paritytech/polkadot-sdk/issues/515
Fixes https://github.com/paritytech/polkadot-sdk/issues/554
Fixes https://github.com/paritytech/polkadot-sdk/issues/556
---
These changes are transferred from
https://github.com/paritytech/substrate/pull/14197 but there are no
functional changes compared to that PR
---------
Co-authored-by: Dmitry Markin <dmitry@markin.tech>
Co-authored-by: Alexandru Vasile <60601340+lexnv@users.noreply.github.com>
* Make peer evictions less aggressive
The original implementation of peer eviction prioritized aliveness over
connection stability which made the peer count unstable for some users.
As this may cause discomfort or infrastructure alerts if stability is
tracked, adjust the eviction to be less aggressive by only evicting
peers when the node has fully stalled. This causes the node to have some
peers who are inactive and won't send any block announcements.
These nodes are removed if the local node is able to receive at least
one block announcement from one of its peers as the inactivity of the
substream is detected when a notification is sent.
If the node won't send or receive any block annoucements for 30 seconds,
it's considered stalled and it will evict all peers,
causing `ProtocolController` to accept and establish connections from new
peers.
* Update client/network/sync/src/engine.rs
Co-authored-by: Dmitry Markin <dmitry@markin.tech>
* Track last send and received notification simultaneously
---------
Co-authored-by: Dmitry Markin <dmitry@markin.tech>
Co-authored-by: parity-processbot <>
If authority discovery is not enabled, `Overseer` is not enabled,
meaning `NetworkBridge` is not started. Validation/collation protocols
are, however, enabled even if the `NetworkBridge` is not started.
Currently this results in normal Polkadot full nodes advertising these
protocols, accepting inbound substreams and even establishing outbound
substreams for the validation protocol. Since the `NetworkBridge` is
not started and no protocol in Substrate is interested in these
protocol events, the events are relayed to all protocol handlers but
are getting discarded because no installed protocol is interested in them.
Co-authored-by: parity-processbot <>
* Accept only `--in-peers` many inbound full nodes in `SyncingEngine`
Due to full and light nodes being stored in the same set, it's possible
that `SyncingEngine` accepts more than `--in-peers` many inbound full
nodes which leaves some of its outbound slots unoccupied.
`ProtocolController` still tries to occupy these slots by opening
outbound substreams. As these substreams are accepted by the remote peer,
the connection is relayed to `SyncingEngine` which rejects the node
because it's already full. This in turn results in the substream being
inactive and the peer getting evicted.
Fixing this properly would require relocating the light peer slot
allocation away from `ProtocolController` or alternatively moving entire
the substream validation there, both of which are epic refactorings and
not necessarily in line with other goals. As a temporary measure, verify
in `SyncingEngine` that it doesn't accept more than the specified amount
of inbound full peers.
* Fix tests
* Apply review comments
There is a race condition in `NetworkBridgeRx` between registering
incoming peers and handling overseer signals. While the peer connection
was sent to `NetworkBridgeRx` first, sometimes the peer would be added to
`validation/collation_peers` with enough delay that the bridge handled
an overseer signal that was meant do send notifications to the registered
peer and as the peer was not present in the bridge, the notification was
never sent and the test would hang.
When peers are registered to `NetworkBridgeRx` using `connect_peer()`,
wait until they show up in `shared.validation/collation_peers` before
doing anything else.
Co-authored-by: parity-processbot <>
* Don't start evicting peers right after `SyncingEngine` is started
Parachain collators may need to wait to receive a relaychain block before
they can start producing blocks which can cause `SyncingEngine` to
incorrectly evict them.
When `SyncingEngine` is started, wait 2 minutes before the eviction is
activated to give collators a chance to produce a block.
* fix doc
* Use `continue` instead of `break`
* Trigger CI
---------
Co-authored-by: parity-processbot <>
* Prepare `sc-network` for `ProtocolController`/`NotificationService`
The upcoming notification protocol refactoring requires that protocols
are able to communicate with `sc-network` over unique and direct links.
This means that `sc-network` side of the link has to be created before
`sc-network` is initialized and that it is allowed to consume the object
as the receiver half of the link may not implement `Clone`.
Remove request-response and notification protocols from `NetworkConfiguration`
and create a new object that contains the configurations of these protocols
and which is consumable by `sc-network`. This is needed needed because, e.g.,
the receiver half of `NotificationService` is not clonable so `sc-network`
must consume it when it's initializing the protocols in `Notifications`.
Similar principe applies to `PeerStore`/`ProtocolController`: as per current
design, protocols are created before the network so `Protocol` cannot be
the one creating the `PeerStore` object. `FullNetworkConfiguration` will be
used to store the objects that `sc-network` will use to communicate with
protocols and it will also allow protocols to allocate handles so they
can directly communicate with `sc-network`.
* Fixes
* Update client/service/src/builder.rs
Co-authored-by: Dmitry Markin <dmitry@markin.tech>
* Updates
* Doc updates + cargo-fmt
---------
Co-authored-by: Dmitry Markin <dmitry@markin.tech>
* Evict inactive peers from `SyncingEngine`
If both halves of the block announce notification stream have been
inactive for 2 minutes, report the peer and disconnect it, allowing
`SyncingEngine` to free up a slot for some other peer that hopefully
is more active.
This needs to be done because the node may falsely believe it has open
connections to peers because the inbound substream can be closed without
any notification and closed outbound substream is noticed only when node
attempts to write to it which may not happen if the node has nothing to
send.
* zzz
* wip
* Evict peers only when timeout expires
* Use `debug!()`
---------
Co-authored-by: parity-processbot <>
* Keep track of the pending response for each peer individually
When peer disconnects or the syncing is restarted, remove the pending
response so syncing won't start sending duplicate requests/receive stale
responses from disconnected peers.
Before this commit pending responses where stored in `FuturesUnordered`
which made it hard to keep track of pending responses for each individual
peer.
* Update client/network/sync/src/lib.rs
Co-authored-by: Bastian Köcher <git@kchr.de>
* ".git/.scripts/commands/fmt/fmt.sh"
* Apply suggestions from code review
Co-authored-by: Dmitry Markin <dmitry@markin.tech>
Co-authored-by: Sebastian Kunert <skunert49@gmail.com>
* Update client/network/sync/src/lib.rs
---------
Co-authored-by: Bastian Köcher <git@kchr.de>
Co-authored-by: command-bot <>
Co-authored-by: Dmitry Markin <dmitry@markin.tech>
Co-authored-by: Sebastian Kunert <skunert49@gmail.com>
* Poll the substream validation before polling `Notifications`
In tests, it can happen that `Notifications` doesn't produce any events
which causes `poll()` to return `Poll::Pending` and the substream
validation futures won't get polled.
Poll the futures before calling `Notifications` so results for substream
validations are received even if `Notifications` is not producing any
events.
* Remove `pending_messages`
* Remove unused import
* Attempt to relieve pressure on `mpsc_network_worker`
`SyncingEngine` interacting with `NetworkWorker` can put a lot of strain
on the channel if the number of inbound connections is high. This is
because `SyncingEngine` is notified of each inbound substream which it
then can either accept or reject and this causes a lot of message
exchange on the already busy channel.
Use a direct channel pair between `Protocol` and `SyncingEngine`
to exchange notification events. It is a temporary change to alleviate
the problems caused by syncing being an independent protocol and the
fix will be removed once `NotificationService` is implemented.
* Apply review comments
* fixes
* trigger ci
* Fix tests
Verify that both peers have a connection now that the validation goes
through `SyncingEngine`. Depending on how the tasks are scheduled,
one of them might not have the peer registered in `SyncingEngine` at which
point the test won't make any progress because block announcement received
from an unknown peer is discarded.
Move polling of `ChainSync` at the end of the function so that if a block
announcement causes a block request to be sent, that can be sent in the
same call to `SyncingEngine::poll()`.
---------
Co-authored-by: parity-processbot <>
`Protocol` is not a reliable source for the information of connected
peers because it doesn't have real-time information of the actual
connectivity state because it's not resposible for accepting/rejecting
connections and gets that information with delay from `SyncinEngine`.
* Move service tests to `client/network/tests`
These tests depend on `sc-network` and `sc-network-sync` so they should
live outside the crate.
* Move some configs from `sc-network-common` to `sc-network`
* Move `NetworkService` traits to `sc-network`
* Move request-responses to `sc-network`
* Remove more stuff
* Remove rest of configs from `sc-network-common` to `sc-network`
* Remove more stuff
* Fix warnings
* Update client/network/src/request_responses.rs
Co-authored-by: Dmitry Markin <dmitry@markin.tech>
* Fix cargo doc
---------
Co-authored-by: Dmitry Markin <dmitry@markin.tech>
* Move import queue out of `sc-network`
Add supplementary asynchronous API for the import queue which means
it can be run as an independent task and communicated with through
the `ImportQueueService`.
This commit removes removes block and justification imports from
`sc-network` and provides `ChainSync` with a handle to import queue so
it can import blocks and justifications. Polling of the import queue is
moved complete out of `sc-network` and `sc_consensus::Link` is
implemented for `ChainSyncInterfaceHandled` so the import queue
can still influence the syncing process.
* Move stuff to SyncingEngine
* Move `ChainSync` instanation to `SyncingEngine`
Some of the tests have to be rewritten
* Move peer hashmap to `SyncingEngine`
* Let `SyncingEngine` to implement `ChainSyncInterface`
* Introduce `SyncStatusProvider`
* Move `sync_peer_(connected|disconnected)` to `SyncingEngine`
* Implement `SyncEventStream`
Remove `SyncConnected`/`SyncDisconnected` events from
`NetworkEvenStream` and provide those events through
`ChainSyncInterface` instead.
Modify BEEFY/GRANDPA/transactions protocol and `NetworkGossip` to take
`SyncEventStream` object which they listen to for incoming sync peer
events.
* Introduce `ChainSyncInterface`
This interface provides a set of miscellaneous functions that other
subsystems can use to query, for example, the syncing status.
* Move event stream polling to `SyncingEngine`
Subscribe to `NetworkStreamEvent` and poll the incoming notifications
and substream events from `SyncingEngine`.
The code needs refactoring.
* Make `SyncingEngine` into an asynchronous runner
This commits removes the last hard dependency of syncing from
`sc-network` meaning the protocol now lives completely outside of
`sc-network`, ignoring the hardcoded peerset entry which will be
addressed in the future.
Code needs a lot of refactoring.
* Fix warnings
* Code refactoring
* Use `SyncingService` for BEEFY
* Use `SyncingService` for GRANDPA
* Remove call delegation from `NetworkService`
* Remove `ChainSyncService`
* Remove `ChainSync` service tests
They were written for the sole purpose of verifying that `NetworWorker`
continues to function while the calls are being dispatched to
`ChainSync`.
* Refactor code
* Refactor code
* Update client/finality-grandpa/src/communication/tests.rs
Co-authored-by: Anton <anton.kalyaev@gmail.com>
* Fix warnings
* Apply review comments
* Fix docs
* Fix test
* cargo-fmt
* Update client/network/sync/src/engine.rs
Co-authored-by: Anton <anton.kalyaev@gmail.com>
* Update client/network/sync/src/engine.rs
Co-authored-by: Anton <anton.kalyaev@gmail.com>
* Add missing docs
* Refactor code
---------
Co-authored-by: Anton <anton.kalyaev@gmail.com>
* Decrease expected peer count in zombinenet tests
Recent change in the in/out peer ratio caused some of the zombienet
tests to fail. Adjust the number of expected peers to reflect the new
ratio to make tests pass again.
* Increase `--out-peers` to 15 for paritydb test
* Don't check the number of connected peers
Establish fewer outbound connections in an attempt to allow publicly
available nodes to accept more full nodes.
Maintain the overall number of connections node should establish.
* Move import queue out of `sc-network`
Add supplementary asynchronous API for the import queue which means
it can be run as an independent task and communicated with through
the `ImportQueueService`.
This commit removes removes block and justification imports from
`sc-network` and provides `ChainSync` with a handle to import queue so
it can import blocks and justifications. Polling of the import queue is
moved complete out of `sc-network` and `sc_consensus::Link` is
implemented for `ChainSyncInterfaceHandled` so the import queue
can still influence the syncing process.
* Fix tests
* Apply review comments
* Apply suggestions from code review
Co-authored-by: Bastian Köcher <git@kchr.de>
* Update client/network/sync/src/lib.rs
Co-authored-by: Bastian Köcher <git@kchr.de>
Co-authored-by: Bastian Köcher <git@kchr.de>
* Do not update peer information if ancestor search is in progress
If block announcement is received from a peer while ancestor search
for that same peer is still in progress, do not update the peer's best
hash and best number as that causes the ancestor search to yield
different information from what was expected and can cause, for example,
a fork of lower height not be be downloaded.
* Block until peers are in sync
Introduce a middleware called `NetworkServiceProvider` which the
`ChainSync` can use to communicate with `NetworkService`. `ChainSync` is
given a `NetworkServiceHandle` which it uses to call `NetworkServiceProvider`
which then dispatches the calls to `NetworkService` on behalf of `ChainSync`.
This change will allow `ChainSync` to disconnect and report peers and
in the future it'll be possible to send requests and notifications
through the `NetworkServiceProvider`.
`NetworkServiceProvider` is needed only until the `ChainSync` object
has been removed from `Protocol`. After that, a normal `NetworkService`
handle can be passed onto `ChainSync` and these changes can be
deprecated.
Co-authored-by: parity-processbot <>
* Refactor service tests in `sc-network`
Create a separate directory for the tests and move common network
instantion related code to `mod.rs` from where it can be used by both
service and chainsync tests.
Use the builder pattern when creating the `TestNetwork` object to reduce
code duplication between the test files.
* Update client/network/src/service/tests/mod.rs
Co-authored-by: Dmitrii Markin <dmitry@markin.tech>
Co-authored-by: Dmitrii Markin <dmitry@markin.tech>
Co-authored-by: parity-processbot <>
* Introduce `ChainSyncInterface`
`ChainSyncInterface` provides an asynchronous interface for other
subsystems to submit calls to `ChainSync`. This allows `NetworkService`
to delegate calls to `ChainSync` while still providing the same API
for other subsystems (for now). This makes it possible to move the
syncing code in piecemeal fashion out of `protocol.rs` as the calls
are just forwarded to `ChainSync`.
* Apply review comments
* Fix tests
* Introduce mockable `ChainSync` object for testing
`mockall` allows to mock `ChainSync` and to verify that the calls made
to `ChaiSync` are firstly executed at all, that they're executed in
correct order and with correct parameters.
This allows to verify, e.g., that delegating calls directly to
`ChainSync` from `NetworkService` still calls the correct functions with
correct arguments even if `Protocol` middleman is removed.
* Add Cargo.lock
* Fix tests
* Update client/network/Cargo.toml
Co-authored-by: Bastian Köcher <git@kchr.de>
* Update Cargo.lock
* Fix clippy and documentation
Co-authored-by: Bastian Köcher <git@kchr.de>
Co-authored-by: parity-processbot <>
Sometimes `NotificationStreamOpenened` would be received for the
other protocol before `SyncConnected` was received so when the
connection was closed, an incorrect event was read from the event
stream.
* Introduce `sc-network-bitswap`
Move the bitswap protocol out of `sc-network` to its own crate.
* Improve test coverage
* Remove mention of bitswap from `sc-network`
* Fix documentation
* Fix clippy and remove the void dependency
* Remove unneeded trait bound and bump prost to 0.11