This cleans up the `Router` code and adds a bunch of tests. These sorts of systems are a real pain to test, since they have a bunch of asynchronous goroutines living their own lives, so the test coverage is decent but not fantastic. Luckily we've been able to move all of the complex peer management and transport logic outside of the router, as synchronous components that are much easier to test, so the core router logic is fairly small and simple.
This also provides some initial test tooling in `p2p/p2ptest` that automatically sets up in-memory networks and channels for use in integration tests. It also includes channel-oriented test asserters in `p2p/p2ptest/require.go`, but these have primarily been written for router testing and should probably be adapted or extended for reactor testing.
This renames `PeerAddress` to `NodeAddress`, moves it and `NodeID` into a separate file `address.go`, adds tests for them, and fixes a bunch of bugs and inconsistencies.
This revises the new P2P `Transport` interface and does some preliminary code cleanups and simplifications.
The major change here is to add `Connection.Handshake()` for performing node handshakes (once the stream transport API is implemented, this can be done entirely independent of the transport). This moves most of the handshaking logic into the `Router`, such as prevention of head-of-line blocking, validation of peer's `NodeInfo`, controlling timeouts, and so on. This significantly simplifies transports, completely removes the need for internal goroutines, and shares common logic across all transports. This also allows varying the handshake `NodeInfo` across peers, e.g. to vary `ListenAddr`. Similarly, connection filtering is also moved into the switch/router so that it can be shared between transports.
Fixes#5981, which was caused by changes in Router behavior after the introduction of the peer manager, leading to a race condition that could halt the test.
This is a temporary measure, I'll start tightening up the new P2P core tomorrow and write "real" tests with better test infrastructure.
This patches over a test data race where the logger would try to read struct internals via `reflect` while these were concurrently modified (specifically `MemoryTransport.closeOnce`).
This test occasionally fails because the peer is already stopped. It is unclear to me exactly what this test is supposed to do, since calling `FlushStop()` will stop the peer, but the test asserts that the peer shouldn't have been stopped by `FlushStop()` since calling `Stop()` afterwards will error in that case.
The current PEX reactor will be removed in the new P2P stack anyway.
Fixes#5998. Sometimes the connection returns "use of closed network connection" instead, so for now we just accept any error. The switch is not long for this world anyway.
This changes the new prototype PEX reactor to resolve peer address URLs into IP/port PEX addresses itself. Branched off of #5974.
I've spent some time thinking about address handling in the P2P stack. We currently use `PeerAddress` URLs everywhere, except for two places: when dialing a peer, and when exchanging addresses via PEX. We had two options:
1. Resolve addresses to endpoints inside `PeerManager`. This would introduce a lot of added complexity: we would have to track connection statistics per endpoint, have goroutines that asynchronously resolve and refresh these endpoints, deal with resolve scheduling before dialing (which is trickier than it sounds since it involves multiple goroutines in the peer manager and router and messes with peer rating order), handle IP address visibility issues, and so on.
2. Resolve addresses to endpoints (IP/port) only where they're used: when dialing, and in PEX. Everywhere else we use URLs.
I went with 2, because this significantly simplifies the handling of hostname resolution, and because I really think the PEX reactor should migrate to exchanging URLs instead of IP/port numbers anyway -- this allows operators to use DNS names for validators (and can easily migrate them to new IPs and/or load balance requests), and also allows different protocols (e.g. QUIC and `MemoryTransport`). Happy to discuss this.
Fixes#5899 by renaming a bunch of P2P Protobuf entities (while maintaining wire compatibility):
* `Message` to `PexMessage` (as it's only used for PEX messages).
* `PexAddrs` to `PexResponse`.
* `PexResponse.Addrs` to `PexResponse.Addresses`.
* `NetAddress` to `PexAddress` (as it's only used by PEX).
Adds a naïve `PeerManager.Advertise()` method that the new PEX reactor can use to fetch addresses to advertise, as well as some other `FIXME`s on address advertisement.
Follow-up from #5947, branched off of #5954.
This simplifies the upgrade logic by adding explicit eviction requests, which can also be useful for other use-cases (e.g. if we need to ban a peer that's misbehaving). Changes:
* Add `evict` map which queues up peers to explicitly evict.
* `upgrading` now only tracks peers that we're upgrading via dialing (`DialNext` → `Dialed`/`DialFailed`).
* `Dialed` will unmark `upgrading`, and queue `evict` if still beyond capacity.
* `Accepted` will pick a random lower-scored peer to upgrade to, if appropriate, and doesn't care about `upgrading` (the dial will fail later, since it's already connected).
* `EvictNext` will return a peer scheduled in `evict` if any, otherwise if beyond capacity just evict the lowest-scored peer.
This limits all of the `upgrading` logic to `DialNext`, `Dialed`, and `DialFailed`, making it much simplier, and it should generally do the right thing in all cases I can think of.
This improves the `peerStore` prototype by e.g.:
* Using a database with Protobuf for persistence, but also keeping full peer set in memory for performance.
* Simplifying the API, by taking/returning struct copies for safety, and removing errors for in-memory operations.
* Caching the ranked peer set, as a temporary solution until a better data structure is implemented.
* Adding `PeerManagerOptions.MaxPeers` and pruning the peer store (based on rank) when it's full.
* Rewriting `PeerAddress` to be independent of `url.URL`, normalizing it and tightening semantics.
See #5936 and #5938 for background.
The plan was initially to have `DialNext()` and `EvictNext()` return a channel. However, implementing this became unnecessarily complicated and error-prone. As an example, the channel would be both consumed and populated (via method calls) by the same driving method (e.g. `Router.dialPeers()`) which could easily cause deadlocks where a method call blocked while sending on the channel that the caller itself was responsible for consuming (but couldn't since it was busy making the method call). It would also require a set of goroutines in the peer manager that would interact with the goroutines in the router in non-obvious ways, and fully populating the channel on startup could cause deadlocks with other startup tasks. Several issues like these made the solution hard to reason about.
I therefore simply made `DialNext()` and `EvictNext()` block until the next peer was available, using internal triggers to wake these methods up in a non-blocking fashion when any relevant state changes occurred. This proved much simpler to reason about, since there are no goroutines in the peer manager (except for trivial retry timers), nor any blocking channel sends, and it instead relies entirely on the existing goroutine structure of the router for concurrency. This also happens to be the same pattern used by the `Transport.Accept()` API, following Go stdlib conventions, so all router goroutines end up using a consistent pattern as well.
Fixes#5941.
Not entirely sure that this will fix the problem (couldn't reproduce), but in any case this is an artifact of a hack in the P2P transport refactor to make it work with the legacy P2P stack, and will be removed when the refactor is done anyway.
This improves the prototype peer manager by:
* Exporting `PeerManager`, making it accessible by e.g. reactors.
* Replacing `Router.SubscribePeerUpdates()` with `PeerManager.Subscribe()`.
* Tracking address/peer connection statistics, and retrying dial failures with exponential backoff.
* Prioritizing peers, with persistent peers configuration.
* Limiting simultaneous connections.
* Evicting peers and upgrading to higher-priority peers.
* Tracking peer heights, as a workaround for legacy shared peer state APIs.
This is getting to a point where we need to determine precise semantics and implement tests, so we should figure out whether it's a reasonable abstraction that we want to use. The main questions are around the API model (i.e. synchronous method calls with the router polling the manager, vs. an event-driven model using channels, vs. the peer manager calling methods on the router to connect/disconnect peers), and who should have the responsibility of managing actual connections (currently the router, while the manager only tracks peer state).
This adds a prototype peer lifecycle manager, `peerManager`, which stores peer data in an internal `peerStore`. The overall idea here is to have methods for peer lifecycle events which exchange a very narrow subset of peer data, and to keep all of the peer metadata (i.e. the `peerInfo` struct) internal, to decouple this from the router and simplify concurrency control. See `peerManager` GoDoc for more information.
The router is still responsible for actually dialing and accepting peer connections, and routing messages across them, but the peer manager is responsible for determining which peers to dial next, preventing multiple connections being established for the same peer (e.g. both inbound and outbound), and making sure we don't dial the same peer several times in parallel. Later it will also track retries and exponential backoff, as well as peer and address quality. It also assumes responsibility for peer updates subscriptions.
It's a bit unclear to me whether we want the peer manager to take on the responsibility of actually dialing and accepting connections as well, or if it should only be tracking peer state for the router while the router is responsible for all transport concerns. Let's revisit this later.
Early but functional prototype of the new `p2p.Router`, see its GoDoc comment for details on how it works. Expect much of this logic to change and improve as we evolve the new P2P stack.
There is a simple test that sets up an in-memory network of four routers with reactors and passes messages between them, but otherwise no exhaustive tests since this is very much a work-in-progress.
The MTU (Maximum Transmission Unit) for Ethernet is 1500 bytes.
The IP header and the TCP header take up 20 bytes each at least (unless
optional header fields are used) and thus the max for (non-Jumbo frame)
Ethernet is 1500 - 20 -20 = 1460
Source: https://stackoverflow.com/a/3074427/820520
While debugging the mempool issue (#5796), I've noticed we're spending
quite a bit of time encoding blobs of data, which never get printed! The
reason is filtering occurs on the level below, so Go runtime rightfully
evaluates function arguments.
I think it's okay to not format raw bytes.
The `NodeInfo` interface does not appear to serve any purpose at all, so I removed it and renamed the `DefaultNodeInfo` struct to `NodeInfo` (including the Protobuf representations). Let me know if this is actually needed for anything.
Only the Protobuf rename is listed in the changelog, since we do not officially support API stability of the `p2p` package (according to `README.md`). The on-wire protocol remains compatible.
This implements a new `Transport` interface and related types for the P2P refactor in #5670. Previously, `conn.MConnection` was very tightly coupled to the `Peer` implementation -- in order to allow alternative non-multiplexed transports (e.g. QUIC), MConnection has now been moved below the `Transport` interface, as `MConnTransport`, and decoupled from the peer. Since the `p2p` package is not covered by our Go API stability, this is not considered a breaking change, and not listed in the changelog.
The initial approach was to implement the new interface in its final form (which also involved possible protocol changes, see https://github.com/tendermint/spec/pull/227). However, it turned out that this would require a large amount of changes to existing P2P code because of the previous tight coupling between `Peer` and `MConnection` and the reliance on subtleties in the MConnection behavior. Instead, I have broadened the `Transport` interface to expose much of the existing MConnection interface, preserved much of the existing MConnection logic and behavior in the transport implementation, and tried to make as few changes to the rest of the P2P stack as possible. We will instead reduce this interface gradually as we refactor other parts of the P2P stack.
The low-level transport code and protocol (e.g. MConnection, SecretConnection and so on) has not been significantly changed, and refactoring this is not a priority until we come up with a plan for QUIC adoption, as we may end up discarding the MConnection code entirely.
There are no tests of the new `MConnTransport`, as this code is likely to evolve as we proceed with the P2P refactor, but tests should be added before a final release. The E2E tests are sufficient for basic validation in the meanwhile.
closes: #5770closes: #5769
also, include node ID in the output (#5769) and modify NodeKey to use
value semantics (it makes perfect sense for NodeKey to not be a
pointer).
After a reactor has failed to parse an incoming message, it shouldn't output the "bad" data into the logs, as that data is unfiltered and could have anything in it. (We also don't think this information is helpful to have in the logs anyways.)
*testing.T.TempDir() causes test cases to fail when
it is unable to remove the temporary directory once
the test case execution terminates. This seems to
happen often with pex reactor test cases.
Marko
Anton Kaliaev
Callum Waters
Erik Grinaker
Aleksandr Bezobchuk
Tess Rinearson
Alessio Treglia