summaryrefslogtreecommitdiffstats
path: root/src
diff options
context:
space:
mode:
Diffstat (limited to 'src')
-rw-r--r--src/node.rs15
-rw-r--r--src/state/host.rs26
-rw-r--r--src/state/mod.rs2
-rw-r--r--src/state/peer.rs90
-rw-r--r--src/state/timing.rs14
-rw-r--r--src/swarm/bootstrap.rs1
-rw-r--r--src/swarm/peer/inbound.rs11
-rw-r--r--src/swarm/peer/keepalive.rs98
-rw-r--r--src/swarm/peer/mod.rs5
-rw-r--r--src/swarm/peer/worker.rs35
-rw-r--r--src/swarm/rtt.rs108
-rw-r--r--src/transport/firewall.rs107
-rw-r--r--src/transport/identity.rs59
-rw-r--r--src/transport/mod.rs7
-rw-r--r--src/transport/paths.rs188
-rw-r--r--src/transport/traffic.rs8
16 files changed, 536 insertions, 238 deletions
diff --git a/src/node.rs b/src/node.rs
index 8e4ace7..97dec79 100644
--- a/src/node.rs
+++ b/src/node.rs
@@ -23,7 +23,7 @@ use crate::swarm::state_file::StateFile;
use crate::swarm::sync::SyncEvent;
use crate::swarm::{self, peer, sync};
use crate::transport::Transport;
-use crate::transport::firewall::{Firewall, FirewallTransport};
+use crate::transport::firewall::Firewall;
use crate::transport::identity::{IdentityConfig, IdentityTransport};
use crate::transport::traffic::{CountingTransport, Traffic};
@@ -77,6 +77,7 @@ impl Node {
config.transport,
identity,
Arc::clone(&traffic),
+ Some(Arc::clone(&firewall)),
)) as Arc<dyn Transport>,
None => config.transport,
};
@@ -91,13 +92,11 @@ impl Node {
discovered_swarm_id: None,
})),
peer_states: Arc::new(RwLock::new(HashMap::new())),
- // Counted outside the firewall: what it drops outbound counts as sent (it
- // plays a lossy network), what it drops inbound never arrived. Both above
- // the identity layer, so by node ID.
- transport: Arc::new(CountingTransport::new(
- Arc::new(FirewallTransport::new(transport, Arc::clone(&firewall))),
- Arc::clone(&traffic),
- )),
+ // Counted above the firewall (in the paths layer, under the identity layer):
+ // what it drops outbound counts as sent (it plays a lossy network), what it
+ // drops inbound never arrived. Without the identity layer (tests) there is
+ // no firewall.
+ transport: Arc::new(CountingTransport::new(transport, Arc::clone(&traffic))),
firewall,
traffic,
stats: Default::default(),
diff --git a/src/state/host.rs b/src/state/host.rs
index 441139e..6eab803 100644
--- a/src/state/host.rs
+++ b/src/state/host.rs
@@ -161,14 +161,24 @@ impl Host for AppState {
fn resume(&self) -> impl Future<Output = ()> + Send {
let state = self.clone();
async move {
- let mut lifecycle = state.lifecycle.write().await;
- if let NodeLifecycle::Joining {
- discovered_swarm_id: Some(swarm_id),
- } = &*lifecycle
- {
- let swarm_id = swarm_id.clone();
- debug!(swarm_id = %swarm_id, "Node resumed");
- *lifecycle = NodeLifecycle::Ready { swarm_id };
+ let resumed = {
+ let mut lifecycle = state.lifecycle.write().await;
+ let swarm_id = match &*lifecycle {
+ NodeLifecycle::Joining {
+ discovered_swarm_id: Some(swarm_id),
+ } => Some(swarm_id.clone()),
+ _ => None,
+ };
+ if let Some(swarm_id) = &swarm_id {
+ debug!(swarm_id = %swarm_id, "Node resumed");
+ *lifecycle = NodeLifecycle::Ready {
+ swarm_id: swarm_id.clone(),
+ };
+ }
+ swarm_id.is_some()
+ };
+ if resumed {
+ crate::swarm::peer::announce(&state);
}
}
}
diff --git a/src/state/mod.rs b/src/state/mod.rs
index 20f92c5..d8291a7 100644
--- a/src/state/mod.rs
+++ b/src/state/mod.rs
@@ -24,7 +24,7 @@ mod read;
mod session;
mod timing;
-pub(crate) use peer::{HelloState, Incidents, Liveness, PeerState, Standing};
+pub(crate) use peer::{HelloState, Incidents, Keepalive, Liveness, PeerState, Standing};
pub(crate) use read::{ReadSource, StaleView};
pub(crate) use session::SyncSession;
pub(crate) use timing::Timing;
diff --git a/src/state/peer.rs b/src/state/peer.rs
index c9356ea..1f08767 100644
--- a/src/state/peer.rs
+++ b/src/state/peer.rs
@@ -67,8 +67,8 @@ const HELLO_RETRY_MAX: Duration = Duration::from_secs(30);
#[derive(Clone)]
pub(crate) struct PeerState {
- /// Whether we lead the pair (`swarm::pair`): we send the keepalive STATUSes. Fixed for
- /// the pair, so computed once.
+ /// Whether we lead the pair (`swarm::pair`): we send the keepalives. Fixed for the
+ /// pair, so computed once.
pub(crate) leads: bool,
pub(crate) standing: Standing,
@@ -85,6 +85,9 @@ pub(crate) struct PeerState {
pub(crate) liveness: Liveness,
pub(crate) added_at: Instant,
pub(crate) last_seen: Option<Instant>,
+ /// Our last keepalive to it (a PING or a STATUS).
+ pub(crate) last_keepalive_at: Option<Instant>,
+ /// Our last STATUS to it, for whatever reason (`status_refresh`).
pub(crate) last_status_at: Option<Instant>,
/// Last time a sync message from this peer was applied (survives resets: the
@@ -133,7 +136,8 @@ pub(crate) struct PeerState {
}
/// Trouble with a peer, counted since it was first listed. Lost keepalives and refused
-/// sends are in the traffic counters (`transport::traffic`): every STATUS_REPLY answers a STATUS.
+/// sends are in the traffic counters (`transport::traffic`): every PONG answers a PING,
+/// every STATUS_REPLY a STATUS.
#[derive(Debug, Default, Clone, Copy, serde::Serialize)]
#[serde(rename_all = "camelCase")]
pub(crate) struct Incidents {
@@ -157,6 +161,15 @@ impl Incidents {
}
}
+/// What a keepalive due is (`PeerState::take_keepalive_due`).
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(crate) enum Keepalive {
+ /// A PING: the peer's transport answers it with a PONG.
+ Ping,
+ /// A STATUS: our state, answered with the peer's.
+ Status,
+}
+
/// Unanswered HELLOs after which a peer counts as unreachable (1 + 2 + 4 s apart).
const UNREACHABLE_AFTER_HELLOS: u32 = 3;
@@ -172,6 +185,7 @@ impl PeerState {
liveness: Liveness::Active,
added_at: Instant::now(),
last_seen: None,
+ last_keepalive_at: None,
last_status_at: None,
last_sync_rx: None,
last_resync_at: None,
@@ -287,16 +301,21 @@ impl PeerState {
true
}
- /// True (and records the time) if an established, quiet peer needs a keepalive.
+ /// The keepalive an established, quiet peer needs, if any (and records it): a PING,
+ /// or a STATUS if we sent it none for `status_refresh`.
///
- /// One STATUS and its STATUS_REPLY tell both sides that the other is alive, so only one side of
- /// a pair needs to send them: the one leading the pair (`leads`). Each of its STATUSes
- /// makes the other side's peer look recently heard from, so the other side never
- /// sends while it does. If they stop, the other side sends once it has heard
- /// nothing for half an interval more. Without that, both sides would get due at about
- /// the same time (each one's quiet time starts with the same exchange) and their
- /// STATUSes would often cross.
- pub(crate) fn take_keepalive_due(&mut self, timing: &Timing, now: Instant) -> bool {
+ /// A PING and its PONG (or a STATUS and its STATUS_REPLY) tell both sides that the
+ /// other is alive, so only one side of a pair needs to send them: the one leading the
+ /// pair (`leads`). Each of its keepalives makes the other side's peer look recently
+ /// heard from, so the other side never sends while it does. If they stop, the other
+ /// side sends once it has heard nothing for half an interval more. Without that, both
+ /// sides would get due at about the same time (each one's quiet time starts with the
+ /// same exchange) and their keepalives would often cross.
+ pub(crate) fn take_keepalive_due(
+ &mut self,
+ timing: &Timing,
+ now: Instant,
+ ) -> Option<Keepalive> {
let quiet_enough = if self.leads {
timing.keepalive
} else {
@@ -307,16 +326,24 @@ impl PeerState {
&& now.saturating_duration_since(self.added_at) >= timing.keepalive
&& self.quiet_for(now) >= quiet_enough
&& self
- .last_status_at
+ .last_keepalive_at
.is_none_or(|t| now.saturating_duration_since(t) >= timing.keepalive);
-
- if due {
- self.last_status_at = Some(now);
+ if !due {
+ return None;
}
- due
+ self.last_keepalive_at = Some(now);
+ let status = self
+ .last_status_at
+ .is_none_or(|t| now.saturating_duration_since(t) >= timing.status_refresh);
+ Some(if status {
+ Keepalive::Status
+ } else {
+ Keepalive::Ping
+ })
}
- /// Self-healing check on STATUS/STATUS_REPLY. Returns true if a re-handshake should start.
+ /// Self-healing check on a STATUS or STATUS_REPLY. Returns true if a re-handshake
+ /// should start.
pub(crate) fn needs_heal(
&mut self,
local: &transport::HeadsSummary,
@@ -378,6 +405,33 @@ impl PeerState {
mod tests {
use super::*;
+ /// Keepalives are PINGs, every `status_refresh` a STATUS; only when quiet.
+ #[test]
+ fn keepalives_are_pings_and_now_and_then_a_status() {
+ let timing = Timing::from_keepalive(Duration::from_secs(60), Duration::from_secs(300));
+ let mut peer = PeerState::new(true, &timing);
+ peer.hello_state = HelloState::Established;
+ let start = peer.added_at;
+ assert_eq!(peer.take_keepalive_due(&timing, start), None); // Too new.
+ let mut kinds = Vec::new();
+ for interval in 1..=6 {
+ let now = start + timing.keepalive * interval;
+ let due = peer.take_keepalive_due(&timing, now);
+ if due == Some(Keepalive::Status) {
+ peer.last_status_at = Some(now);
+ }
+ kinds.extend(due);
+ // Not again within the interval.
+ assert_eq!(peer.take_keepalive_due(&timing, now), None);
+ }
+ use Keepalive::{Ping, Status};
+ assert_eq!(kinds, [Status, Ping, Ping, Status, Ping, Ping]);
+ // Heard from just now: nothing due.
+ let now = start + timing.keepalive * 7;
+ peer.mark_seen(now);
+ assert_eq!(peer.take_keepalive_due(&timing, now), None);
+ }
+
/// R41: unreachability lasts until the peer is heard from, not for a set time, and
/// probes back off meanwhile.
#[test]
diff --git a/src/state/timing.rs b/src/state/timing.rs
index 6ce5bef..9055ed6 100644
--- a/src/state/timing.rs
+++ b/src/state/timing.rs
@@ -18,6 +18,10 @@ const PROBE_MAX_FACTOR: u32 = 6;
#[derive(Clone, Copy, Debug)]
pub(crate) struct Timing {
pub(crate) keepalive: Duration,
+ /// How often a quiet pair's keepalive is a STATUS rather than a PING
+ /// (`PeerState::take_keepalive_due`): its state (heads, readiness, epoch) for the heal
+ /// and against lost announcements. Changes go out as they happen anyway.
+ pub(crate) status_refresh: Duration,
pub(crate) dormant_after: Duration,
pub(crate) resync_after: Duration,
pub(crate) dormant_probe: Duration,
@@ -43,8 +47,9 @@ impl Timing {
///
/// Each node keeps about `k` neighbours (`--neighbours`) whatever the swarm's size,
/// and only one side of a pair sends them, so idle, a node sends about `k` messages and
- /// receives about `k` (STATUSes and STATUS_REPLYs) per keepalive interval (more on a stable
- /// member others prefer). The interval trades that against how soon a silent peer is
+ /// receives about `k` per keepalive interval (more on a stable member others prefer):
+ /// PINGs and PONGs (a few bytes), every third a STATUS and STATUS_REPLY (heads,
+ /// readiness, epoch) instead. The interval trades that against how soon a silent peer is
/// noticed (`dormant_after`, 4 intervals) and a gap closes (within an interval).
/// Recommended:
///
@@ -68,7 +73,10 @@ impl Timing {
) -> Self {
Self {
keepalive,
- dormant_after: keepalive * 4, // dormant_after >= 4* keepalive to tolerate three lost STATUSes in a row
+ // Every third keepalive; with `resync_after` at two keepalives, a heal needs
+ // two STATUSes that differ: within about two refreshes.
+ status_refresh: keepalive * 3,
+ dormant_after: keepalive * 4, // dormant_after >= 4* keepalive to tolerate three lost keepalives in a row
// >= 2 keepalives, so a second exchange confirms the difference; and >= twice
// the first sync reply timeout, so a lost message's sender resends it first.
resync_after: (keepalive * 2).max(sync_reply_timeout * 2),
diff --git a/src/swarm/bootstrap.rs b/src/swarm/bootstrap.rs
index 7c4c444..a9ab1b2 100644
--- a/src/swarm/bootstrap.rs
+++ b/src/swarm/bootstrap.rs
@@ -119,6 +119,7 @@ pub(crate) async fn bootstrap(state: &AppState, seeds: &[String], start: Start)
if state.initial_sync_done.load(Ordering::Acquire) && !state.remake.is_rejoining() {
state.mark_ready(swarm_id.clone()).await?;
info!(swarm_id = %swarm_id, "Node is ready");
+ crate::swarm::peer::announce(state);
return Ok(());
}
}
diff --git a/src/swarm/peer/inbound.rs b/src/swarm/peer/inbound.rs
index 92a1a5e..0eaefb9 100644
--- a/src/swarm/peer/inbound.rs
+++ b/src/swarm/peer/inbound.rs
@@ -11,7 +11,7 @@ use super::epoch::observe_peer_epoch;
use super::handshake::{
discover_peers, ensure_handshake, establish_peer, peer_swarm, remember_peer, send_hello_reply,
};
-use super::keepalive::{heal_if_diverged, local_status, note_status};
+use super::keepalive::{heal_if_diverged, local_status, note_alive, note_status};
use super::lanes::SyncLanes;
use crate::state::AppState;
use crate::swarm::sync::SyncEvent;
@@ -30,6 +30,11 @@ pub(crate) async fn handle_message(
debug!("Ignoring message from self");
return Ok(());
}
+ // PING and PONG only say the peer is alive (`keepalive::note_alive`).
+ if matches!(message, ParsedMessage::Ping(_) | ParsedMessage::Pong(_)) {
+ note_alive(state, &peer, &message).await;
+ return Ok(());
+ }
// WHO and ME end at the identity layer (`transport::identity`).
let Some((epoch, epoch_seq)) = message.epoch_stamp() else {
return Ok(());
@@ -154,10 +159,6 @@ pub(crate) async fn handle_message(
}
}
transport::ParsedMessage::StatusReply(remote) => {
- // It answers our STATUS: a round-trip sample.
- state
- .lock_rtt()
- .status_reply_received(&peer, Instant::now());
on_status(state, &peer, &remote, reactivated).await;
}
transport::ParsedMessage::Who(_)
diff --git a/src/swarm/peer/keepalive.rs b/src/swarm/peer/keepalive.rs
index 810c377..460b71e 100644
--- a/src/swarm/peer/keepalive.rs
+++ b/src/swarm/peer/keepalive.rs
@@ -1,12 +1,19 @@
-//! Keepalives (STATUS / STATUS_REPLY) and the self-healing re-handshake when they show our
-//! documents differ while sync is idle.
+//! Keepalives and the self-healing re-handshake.
+//!
+//! A quiet pair's keepalive (`PeerState::take_keepalive_due`) is a PING, answered with a
+//! PONG by the peer's paths layer (`transport::paths`), over the same path: it tells both
+//! sides the other is alive, and gives the round trip (`swarm::rtt`). Every `Timing::status_refresh` it is a
+//! STATUS instead, answered with a STATUS_REPLY: each side's state (heads, readiness,
+//! epoch), which also go out as they change (relay announcements, the epoch after a
+//! switch, readiness: `announce`). When a STATUS shows our documents differ while sync
+//! is idle, the pair re-handshakes (`heal_if_diverged`).
use std::sync::atomic::Ordering;
use std::time::Duration;
use anyhow::Result;
use tokio::time::Instant;
-use tracing::warn;
+use tracing::{debug, info, warn};
use super::handshake::ensure_handshake;
use crate::state::{AppState, Liveness, Standing};
@@ -14,33 +21,100 @@ use crate::swarm::sync::SyncEvent;
use crate::swarm::topology;
use crate::transport::{self, ParsedMessage};
-/// Sends `peer` a STATUS with our status, noting the time for its round trip.
+/// Sends `peer` a STATUS with our status.
pub(crate) async fn send_status(state: &AppState, peer: &str) -> Result<()> {
let status = local_status(state, peer).await;
- state.lock_rtt().status_sent(peer, Instant::now());
+ if let Some(peer_state) = state.get_peer_state(peer).await {
+ peer_state.lock().await.last_status_at = Some(Instant::now());
+ }
state
.transport
.send_message(peer, &ParsedMessage::Status(status))
.await
}
+/// Sends `peer` a PING, noting it for its round trip.
+pub(crate) async fn send_ping(state: &AppState, peer: &str) -> Result<()> {
+ let nonce = crate::transport::paths::random_nonce();
+ state.lock_rtt().ping_sent(peer, nonce, Instant::now());
+ state
+ .transport
+ .send_message(peer, &ParsedMessage::Ping(nonce))
+ .await
+}
+
+/// A PING or PONG from `peer`: it is alive, and a PONG may answer our PING (its round
+/// trip). A transport without the paths layer (`Transport::answers_pings`: in tests)
+/// leaves answering a PING to us. Only peers we know already: a PING makes none.
+pub(super) async fn note_alive(state: &AppState, peer: &str, message: &ParsedMessage) {
+ let now = Instant::now();
+ match message {
+ ParsedMessage::Ping(nonce) if !state.transport.answers_pings() => {
+ let pong = ParsedMessage::Pong(*nonce);
+ if let Err(error) = state.transport.send_message(peer, &pong).await {
+ debug!(peer = %peer, error = %format!("{error:#}"), "PONG failed");
+ }
+ }
+ ParsedMessage::Pong(nonce) => state.lock_rtt().pong_received(peer, *nonce, now),
+ _ => {}
+ }
+ let Some(peer_state) = state.get_peer_state(peer).await else {
+ return;
+ };
+ let reactivated = {
+ let mut peer_state = peer_state.lock().await;
+ if peer_state.unreachable_since.is_some() {
+ // Reachable again (`mark_seen` clears it).
+ state.topology_changed.notify_one();
+ }
+ peer_state.mark_seen(now)
+ };
+ crate::swarm::source::heard_from(state, peer);
+ if reactivated {
+ info!(peer = %peer, "Dormant peer is active again; re-handshaking");
+ ensure_handshake(state, peer).await;
+ }
+}
+
/// How long shutting down waits for the last STATUSes (`farewell`).
const FAREWELL_TIMEOUT: Duration = Duration::from_secs(2);
-/// At shutdown: one last STATUS to each neighbour, saying we aren't ready. A neighbour
-/// catching up from us picks another source now, instead of once we go dormant. Bounded
-/// in time: our transport may be gone already.
-pub(crate) async fn farewell(state: &AppState) {
- state.stopping.store(true, Ordering::Release);
+/// A STATUS to each neighbour, at most `within` (all sent: none).
+async fn status_to_neighbours(state: &AppState, within: Option<Duration>) {
let mut statuses = tokio::task::JoinSet::new();
for (peer, _) in state.peers_snapshot().await {
if topology::is_neighbour(state, &peer) {
let state = state.clone();
- statuses.spawn(async move { send_status(&state, &peer).await });
+ statuses.spawn(async move {
+ if let Err(error) = send_status(&state, &peer).await {
+ debug!(peer = %peer, error = %format!("{error:#}"), "STATUS failed");
+ }
+ });
}
}
let all_sent = async { while statuses.join_next().await.is_some() {} };
- let _ = tokio::time::timeout(FAREWELL_TIMEOUT, all_sent).await;
+ match within {
+ Some(within) => {
+ let _ = tokio::time::timeout(within, all_sent).await;
+ }
+ None => all_sent.await,
+ }
+}
+
+/// Our readiness changed (we became ready, or ready again after a rejoin): a STATUS
+/// to each neighbour now, rather than with the next refresh. A peer opens sync towards
+/// us only once it knows we are ready.
+pub(crate) fn announce(state: &AppState) {
+ let state = state.clone();
+ tokio::spawn(async move { status_to_neighbours(&state, None).await });
+}
+
+/// At shutdown: one last STATUS to each neighbour, saying we aren't ready. A neighbour
+/// catching up from us picks another source now, instead of once we go dormant. Bounded
+/// in time: our transport may be gone already.
+pub(crate) async fn farewell(state: &AppState) {
+ state.stopping.store(true, Ordering::Release);
+ status_to_neighbours(state, Some(FAREWELL_TIMEOUT)).await;
}
/// Our status for `peer`.
diff --git a/src/swarm/peer/mod.rs b/src/swarm/peer/mod.rs
index 415cddf..326f3a9 100644
--- a/src/swarm/peer/mod.rs
+++ b/src/swarm/peer/mod.rs
@@ -2,7 +2,8 @@
//! swarm-wide forget records, and dispatch of inbound messages.
//!
//! * `handshake`: discovering peers, HELLO / HELLO_REPLY, membership.
-//! * `keepalive`: STATUS / STATUS_REPLY and the self-healing re-handshake.
+//! * `keepalive`: PING / PONG and STATUS / STATUS_REPLY, and the self-healing
+//! re-handshake.
//! * `epoch`: the epochs peers advertise.
//! * `admit`: admission with a swarm key.
//! * `forget`: swarm-wide forget records.
@@ -24,6 +25,6 @@ pub(crate) use handshake::{
apply_neighbour_change, peer_swarm, rehandshake, remember_peer, send_hello_once,
};
pub(crate) use inbound::handle_message;
-pub(crate) use keepalive::{farewell, send_status};
+pub(crate) use keepalive::{announce, farewell, send_status};
pub(crate) use lanes::SyncLanes;
pub(crate) use worker::run_peer_worker;
diff --git a/src/swarm/peer/worker.rs b/src/swarm/peer/worker.rs
index c68e2ea..f3d0e84 100644
--- a/src/swarm/peer/worker.rs
+++ b/src/swarm/peer/worker.rs
@@ -10,9 +10,9 @@ use tracing::{debug, info, warn};
use super::forget::{AppliedForgets, apply_forget_records, forget_strangers};
use super::handshake::retry_hello_if_due;
-use super::keepalive::send_status;
+use super::keepalive::{send_ping, send_status};
use crate::locks::MutexExt;
-use crate::state::AppState;
+use crate::state::{AppState, Keepalive};
use crate::swarm::sync::SyncEvent;
use crate::swarm::topology;
@@ -50,16 +50,20 @@ pub(crate) async fn run_peer_worker(state: AppState) {
// Only neighbours are kept alive and judged; others we only answer.
let neighbour = topology::is_neighbour(&state, peer);
- let (became_dormant, ping_due, unanswered) = {
+ let (became_dormant, keepalive, unanswered) = {
let mut peer_state = peer_state_mutex.lock().await;
let became_dormant =
ready && neighbour && peer_state.check_dormancy(&state.timing, now);
- let ping_due = neighbour && peer_state.take_keepalive_due(&state.timing, now);
+ let keepalive = if neighbour {
+ peer_state.take_keepalive_due(&state.timing, now)
+ } else {
+ None
+ };
if !neighbour {
peer_state.stop_hello_retries();
}
let unanswered = peer_state.take_overdue_sync_reply(now);
- (became_dormant, ping_due, unanswered)
+ (became_dormant, keepalive, unanswered)
};
if let Some(changes) = unanswered {
@@ -80,12 +84,21 @@ pub(crate) async fn run_peer_worker(state: AppState) {
);
}
- let probe_due =
- !ping_due && crate::swarm::source::probe_due(&state, peer, Instant::now());
- if (ping_due || probe_due)
- && let Err(error) = send_status(&state, peer).await
- {
- debug!(peer = %peer, error = %format!("{error:#}"), "Keepalive STATUS failed");
+ // The catch-up source gets STATUSes: we need its readiness, and whether it
+ // syncs with us.
+ let keepalive = match keepalive {
+ None if crate::swarm::source::probe_due(&state, peer, Instant::now()) => {
+ Some(Keepalive::Status)
+ }
+ keepalive => keepalive,
+ };
+ let sent = match keepalive {
+ Some(Keepalive::Ping) => send_ping(&state, peer).await,
+ Some(Keepalive::Status) => send_status(&state, peer).await,
+ None => Ok(()),
+ };
+ if let Err(error) = sent {
+ debug!(peer = %peer, kind = ?keepalive, error = %format!("{error:#}"), "Keepalive failed");
}
if let Err(error) = retry_hello_if_due(&state, peer).await {
diff --git a/src/swarm/rtt.rs b/src/swarm/rtt.rs
index 01d10ee..466fc98 100644
--- a/src/swarm/rtt.rs
+++ b/src/swarm/rtt.rs
@@ -1,9 +1,11 @@
-//! Round-trip times to peers, measured from traffic we send anyway: STATUS → STATUS_REPLY and
-//! HELLO → HELLO_REPLY. No probes of our own.
+//! Round-trip times to peers, measured from traffic we send anyway: the keepalive PING
+//! → PONG (`swarm::peer::keepalive`), and HELLO → HELLO_REPLY.
//!
-//! A sample only counts when exactly one request was outstanding: with two, we can't tell
-//! which one the answer belongs to (an answer to the older one would look too fast).
-//! Samples are smoothed (exponential moving average, weight 1/4) to ride out jitter.
+//! A PONG answers the PING with its nonce, so it is attributed exactly; a PONG to an
+//! older PING (one a newer one replaced) is no sample. A HELLO_REPLY only counts when
+//! exactly one HELLO was outstanding: with two, we can't tell which one the answer
+//! belongs to (an answer to the older one would look too fast). Samples are smoothed
+//! (exponential moving average, weight 1/4) to ride out jitter.
//!
//! What we use is the *effective* round-trip time: the smoothed value, raised to the age
//! of the oldest request still unanswered (the round trip takes at least that long), so
@@ -18,11 +20,20 @@ use tokio::time::Instant;
#[derive(Debug, Default)]
pub(crate) struct RttTable {
- statuses: HashMap<String, Outstanding>,
+ pings: HashMap<String, Ping>,
hellos: HashMap<String, Outstanding>,
smoothed: HashMap<String, Sample>,
}
+/// The PING out to a peer.
+#[derive(Debug, Clone, Copy)]
+struct Ping {
+ nonce: u64,
+ sent: Instant,
+ /// When the oldest PING still unanswered went out (an earlier one, replaced).
+ waiting_since: Instant,
+}
+
#[derive(Debug, Clone, Copy)]
struct Sample {
rtt: Duration,
@@ -36,13 +47,26 @@ struct Outstanding {
}
impl RttTable {
- pub(crate) fn status_sent(&mut self, peer: &str, now: Instant) {
- Self::sent(&mut self.statuses, peer, now);
+ pub(crate) fn ping_sent(&mut self, peer: &str, nonce: u64, now: Instant) {
+ let waiting_since = self
+ .pings
+ .get(peer)
+ .map_or(now, |earlier| earlier.waiting_since);
+ self.pings.insert(
+ peer.to_owned(),
+ Ping {
+ nonce,
+ sent: now,
+ waiting_since,
+ },
+ );
}
- pub(crate) fn status_reply_received(&mut self, peer: &str, now: Instant) {
- if let Some(sample) = Self::answered(&mut self.statuses, peer, now) {
- self.record(peer, sample, now);
+ pub(crate) fn pong_received(&mut self, peer: &str, nonce: u64, now: Instant) {
+ if self.pings.get(peer).is_some_and(|ping| ping.nonce == nonce)
+ && let Some(ping) = self.pings.remove(peer)
+ {
+ self.record(peer, now.saturating_duration_since(ping.sent), now);
}
}
@@ -68,12 +92,15 @@ impl RttTable {
.smoothed
.get(peer)
.filter(|sample| now.saturating_duration_since(sample.at) < fresh_for)?;
- let waiting = [self.statuses.get(peer), self.hellos.get(peer)]
- .into_iter()
- .flatten()
- .map(|outstanding| now.saturating_duration_since(outstanding.sent))
- .max()
- .unwrap_or_default();
+ let waiting = [
+ self.pings.get(peer).map(|ping| ping.waiting_since),
+ self.hellos.get(peer).map(|outstanding| outstanding.sent),
+ ]
+ .into_iter()
+ .flatten()
+ .map(|sent| now.saturating_duration_since(sent))
+ .max()
+ .unwrap_or_default();
Some(sample.rtt.max(waiting))
}
@@ -128,31 +155,47 @@ mod tests {
let mut table = RttTable::default();
let start = Instant::now();
let at = |ms| start + Duration::from_millis(ms);
- table.status_sent("a", start);
- table.status_reply_received("a", at(100));
+ table.ping_sent("a", 1, start);
+ table.pong_received("a", 1, at(100));
assert_eq!(
table.effective("a", at(100), FRESH),
Some(Duration::from_millis(100))
);
- table.status_sent("a", at(100));
- table.status_reply_received("a", at(300));
+ table.ping_sent("a", 2, at(100));
+ table.pong_received("a", 2, at(300));
assert_eq!(
table.effective("a", at(300), FRESH),
Some(Duration::from_millis(125))
);
- // Two STATUSes out: the STATUS_REPLY can't be attributed.
- table.status_sent("a", at(300));
- table.status_sent("a", at(400));
- table.status_reply_received("a", at(450));
+ // A newer PING replaced it: a PONG to the older one is no sample.
+ table.ping_sent("a", 3, at(300));
+ table.ping_sent("a", 4, at(400));
+ table.pong_received("a", 3, at(410));
assert_eq!(
- table.effective("a", at(450), FRESH),
+ table.effective("a", at(410), FRESH),
Some(Duration::from_millis(125))
);
+ table.pong_received("a", 4, at(500));
+ assert_eq!(
+ table.effective("a", at(500), FRESH),
+ Some(Duration::from_micros(118_750)) // (125 * 3 + 100) / 4
+ );
+
+ // Two HELLOs out: the HELLO_REPLY can't be attributed.
+ table.hello_sent("c", start);
+ table.hello_reply_received("c", at(100));
+ table.hello_sent("c", at(100));
+ table.hello_sent("c", at(200));
+ table.hello_reply_received("c", at(210));
+ assert_eq!(
+ table.effective("c", at(210), FRESH),
+ Some(Duration::from_millis(100))
+ );
- // An unsolicited STATUS_REPLY changes nothing.
- table.status_reply_received("b", start);
+ // An unsolicited PONG changes nothing.
+ table.pong_received("b", 9, start);
assert_eq!(table.effective("b", start, FRESH), None);
}
@@ -160,12 +203,13 @@ mod tests {
fn unanswered_requests_raise_it_and_old_samples_expire() {
let mut table = RttTable::default();
let start = Instant::now();
- table.status_sent("a", start);
- table.status_reply_received("a", start + Duration::from_millis(50));
+ table.ping_sent("a", 1, start);
+ table.pong_received("a", 1, start + Duration::from_millis(50));
- // A STATUS unanswered for 40 s: at least 40 s.
+ // PINGs unanswered for 40 s: at least 40 s, from the oldest.
let sent = start + Duration::from_secs(60);
- table.status_sent("a", sent);
+ table.ping_sent("a", 2, sent);
+ table.ping_sent("a", 3, sent + Duration::from_secs(30));
let later = sent + Duration::from_secs(40);
assert_eq!(
table.effective("a", later, FRESH),
diff --git a/src/transport/firewall.rs b/src/transport/firewall.rs
index 846ae1b..7b0ec0d 100644
--- a/src/transport/firewall.rs
+++ b/src/transport/firewall.rs
@@ -1,13 +1,14 @@
+//! The debug firewall (`/debug/firewall`, with `--debug-api`): drops messages from and
+//! to peers, by node ID, to test partitions on a real swarm. It works at the bottom of
+//! the node, in the paths layer (`transport::paths::Gate`), so it drops everything a
+//! partition would: WHO and ME, PINGs and their PONGs, as well as the node's messages.
+
use std::collections::HashSet;
-use std::sync::{Arc, RwLock, RwLockReadGuard};
+use std::sync::{RwLock, RwLockReadGuard};
-use anyhow::Result;
-use async_trait::async_trait;
use serde::{Deserialize, Serialize};
use crate::locks::RwLockExt;
-use crate::transport::EncodedMessage;
-use crate::transport::{MessageHandler, ParsedMessage, Transport};
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub(crate) struct FirewallRules {
@@ -46,99 +47,3 @@ impl Firewall {
self.read().outbound.contains(peer)
}
}
-
-pub(crate) struct FirewallTransport {
- inner: Arc<dyn Transport>,
- firewall: Arc<Firewall>,
-}
-
-impl FirewallTransport {
- pub(crate) fn new(inner: Arc<dyn Transport>, firewall: Arc<Firewall>) -> Self {
- Self { inner, firewall }
- }
-}
-
-#[async_trait]
-impl Transport for FirewallTransport {
- fn local_id(&self) -> &str {
- self.inner.local_id()
- }
-
- fn is_peer_id(&self, id: &str) -> bool {
- self.inner.is_peer_id(id)
- }
-
- fn delivery_limit(&self) -> std::time::Duration {
- self.inner.delivery_limit()
- }
-
- fn local_addresses(&self) -> Vec<String> {
- self.inner.local_addresses()
- }
-
- fn addresses_of(&self, peer: &str) -> Vec<String> {
- self.inner.addresses_of(peer)
- }
-
- fn learn_addresses(&self, peer: &str, addresses: &[String]) {
- self.inner.learn_addresses(peer, addresses)
- }
-
- fn can_reach(&self, peer: &str) -> bool {
- self.inner.can_reach(peer)
- }
-
- fn discovered_seeds(&self) -> Vec<String> {
- self.inner.discovered_seeds()
- }
-
- fn transports(&self) -> Vec<String> {
- self.inner.transports()
- }
-
- fn destination(&self, peer: &str) -> Option<crate::transport::Destination> {
- self.inner.destination(peer)
- }
-
- fn connection(&self, target: &str) -> Option<crate::transport::Connection> {
- self.inner.connection(target)
- }
-
- fn stats(&self) -> crate::transport::stats::TransportStats {
- self.inner.stats()
- }
-
- async fn resolve_seed(&self, seed: &str) -> Result<String> {
- self.inner.resolve_seed(seed).await
- }
-
- fn encode(&self, peer: &str, message: &ParsedMessage) -> Result<EncodedMessage> {
- self.inner.encode(peer, message)
- }
-
- async fn send_encoded(&self, message: &EncodedMessage) -> Result<()> {
- if self.firewall.blocks_outbound(&message.peer) {
- tracing::trace!(peer = %message.peer, "Firewall dropped outbound message");
- return Ok(());
- }
- self.inner.send_encoded(message).await
- }
-
- async fn start_consumer(&self, handler: MessageHandler) -> Result<()> {
- let firewall = Arc::clone(&self.firewall);
-
- self.inner
- .start_consumer(Box::new(move |sender, message, payload_len| {
- if firewall.blocks_inbound(&sender) {
- tracing::trace!(peer = %sender, "Firewall dropped inbound message");
- return Box::pin(async { Ok(()) });
- }
- handler(sender, message, payload_len)
- }))
- .await
- }
-
- async fn wait_for_exit(&self) -> std::io::Result<std::process::ExitStatus> {
- self.inner.wait_for_exit().await
- }
-}
diff --git a/src/transport/identity.rs b/src/transport/identity.rs
index 5857fb0..153dacc 100644
--- a/src/transport/identity.rs
+++ b/src/transport/identity.rs
@@ -37,7 +37,8 @@ use tracing::{debug, info};
use crate::identity::{self, NodeKey};
use crate::locks::MutexExt;
-use crate::transport::paths::{PathHandler, Paths, scheme_of};
+use crate::transport::firewall::Firewall;
+use crate::transport::paths::{Gate, PathHandler, Paths, scheme_of};
use crate::transport::stats::{IdentityCounts, IdentityStats, TransportStats};
use crate::transport::traffic::Traffic;
use crate::transport::{
@@ -96,6 +97,8 @@ pub(crate) struct IdentityTransport {
struct Shared {
paths: Arc<Paths>,
+ /// The debug firewall, by node ID; paths applies it (`Gate`).
+ firewall: Option<Arc<Firewall>>,
key: Arc<NodeKey>,
node_id: String,
/// WHO and ME are counted here, under the address (`traffic`).
@@ -208,27 +211,54 @@ impl Book {
impl IdentityTransport {
/// Over `transport`, whose addresses have `config.scheme`, and `config.more`.
+ /// `firewall`: the debug firewall, if any, which paths applies by address.
pub(crate) fn new(
transport: Arc<dyn Transport>,
config: IdentityConfig,
traffic: Arc<Traffic>,
+ firewall: Option<Arc<Firewall>>,
) -> Self {
let node_id = config.key.id();
- let paths = Arc::new(Paths::new(
- std::iter::once((config.scheme, transport))
- .chain(config.more)
- .collect(),
- Some(Arc::clone(&traffic)),
- ));
- Self {
- shared: Arc::new(Shared {
- paths,
+ let shared = Arc::new_cyclic(|shared: &std::sync::Weak<Shared>| {
+ let gate: std::sync::Weak<dyn Gate> = shared.clone();
+ Shared {
+ paths: Arc::new(Paths::new(
+ std::iter::once((config.scheme, transport))
+ .chain(config.more)
+ .collect(),
+ Some(Arc::clone(&traffic)),
+ firewall.is_some().then_some(gate),
+ )),
+ firewall,
key: config.key,
node_id,
traffic,
book: Mutex::default(),
- }),
- }
+ }
+ });
+ Self { shared }
+ }
+}
+
+/// The firewall's rules are by node ID: an address is blocked if the node proven at it
+/// is. Addresses no node proved itself at aren't (nothing could name them).
+impl Gate for Shared {
+ fn blocks_inbound(&self, address: &str) -> bool {
+ self.blocks(address, Firewall::blocks_inbound)
+ }
+
+ fn blocks_outbound(&self, address: &str) -> bool {
+ self.blocks(address, Firewall::blocks_outbound)
+ }
+}
+
+impl Shared {
+ fn blocks(&self, address: &str, rule: fn(&Firewall, &str) -> bool) -> bool {
+ let Some(firewall) = &self.firewall else {
+ return false;
+ };
+ let node = self.lock().at.get(address).map(|(node, _)| node.clone());
+ node.is_some_and(|node| rule(firewall, &node))
}
}
@@ -461,6 +491,11 @@ impl Transport for IdentityTransport {
})
}
+ /// The paths layer answers them (`Paths::receive`).
+ fn answers_pings(&self) -> bool {
+ true
+ }
+
/// The identity layer's, and each transport's.
fn stats(&self) -> TransportStats {
let mut stats = self.shared.paths.stats();
diff --git a/src/transport/mod.rs b/src/transport/mod.rs
index 147e560..f60a9f6 100644
--- a/src/transport/mod.rs
+++ b/src/transport/mod.rs
@@ -649,6 +649,13 @@ pub(crate) trait Transport: Send + Sync {
None
}
+ /// Whether the transport answers PINGs with PONGs itself (the paths layer,
+ /// `transport::paths`); if not (a bare transport, in tests), the node does. By
+ /// default not.
+ fn answers_pings(&self) -> bool {
+ false
+ }
+
/// Figures about the transport itself (`stats`). By default none.
fn stats(&self) -> stats::TransportStats {
stats::TransportStats::default()
diff --git a/src/transport/paths.rs b/src/transport/paths.rs
index 30aab8d..dda2d87 100644
--- a/src/transport/paths.rs
+++ b/src/transport/paths.rs
@@ -17,16 +17,24 @@
//!
//! Paths measures the addresses of such nodes itself: every `PROBE_EVERY`, each address
//! it had a choice of within `ACTIVE_FOR` gets a PING with a random nonce, which the
-//! other side answers with a PONG at once, over the same path. Answers give each
+//! other side's paths answers at once with a PONG, over the same path. Answers give each
//! path's round-trip time (smoothed); `UNANSWERED_DOWN` PINGs in a row unanswered, or a
//! send that failed within `DOWN_FOR`, make a path down. `rank` puts paths that are up
//! first, the one in use first among them, and the rest by round-trip time; it changes
//! the one in use when another is faster by `SWITCH_GAIN` (a quarter) and at least
//! `SWITCH_MIN` (both measured), so it doesn't flap. Until both are measured, the
-//! configured order decides. PING and PONG end here: the identity layer never sees them.
+//! configured order decides.
+//!
+//! PINGs and PONGs then go up as other messages do: they tell the node the peer is
+//! alive. The node's keepalive is a PING too (`swarm::peer::keepalive`), which paths
+//! sends like any message, and notes as the probe of the path it went over.
+//!
+//! The debug firewall (`transport::firewall`) works here, below everything else, through
+//! a `Gate`: messages from an address it blocks are dropped as they arrive, and sends to
+//! one are dropped silently, as by a lossy network.
use std::collections::HashMap;
-use std::sync::{Arc, Mutex};
+use std::sync::{Arc, Mutex, Weak};
use std::time::Duration;
use anyhow::{Context, Result, anyhow};
@@ -36,7 +44,14 @@ use tracing::{debug, info};
use crate::locks::MutexExt;
use crate::transport::stats::{PathCounts, PathInfo, PathsStats, TransportStats};
use crate::transport::traffic::Traffic;
-use crate::transport::{Connection, EncodedMessage, MessageKind, ParsedMessage, Transport};
+use crate::transport::{Connection, EncodedMessage, ParsedMessage, Transport};
+
+/// Whether to drop what comes from or goes to an address (the debug firewall,
+/// `transport::firewall`; the identity layer knows which node is where).
+pub(crate) trait Gate: Send + Sync {
+ fn blocks_inbound(&self, address: &str) -> bool;
+ fn blocks_outbound(&self, address: &str) -> bool;
+}
/// Called with the sender's address, the message and its `payload_len`.
pub(crate) type PathHandler = dyn Fn(String, ParsedMessage, usize) -> std::pin::Pin<Box<dyn Future<Output = Result<()>> + Send>>
@@ -250,6 +265,16 @@ impl Table {
due
}
+ /// The node sent a PING with `nonce` to `address` (its keepalive): the probe of that
+ /// path, if paths measures it.
+ fn pinged(&mut self, address: &str, nonce: u64, now: Instant) {
+ if let Some(path) = self.paths.get_mut(address) {
+ path.probe = Some((nonce, now));
+ path.last_probe = Some(now);
+ self.counts.probes_sent += 1;
+ }
+ }
+
/// A PONG with `nonce` from `address`: the round trip, if it answers our PING out.
fn pong(&mut self, address: &str, nonce: u64, now: Instant) -> Option<Duration> {
let path = self.paths.get_mut(address)?;
@@ -298,18 +323,21 @@ impl Table {
pub(crate) struct Paths {
/// In order of preference.
links: Vec<Link>,
- /// PINGs and PONGs are counted here, under the address (`traffic`).
+ /// Its own PINGs and PONGs are counted here, under the address (`traffic`).
traffic: Option<Arc<Traffic>>,
+ /// The debug firewall, if any.
+ gate: Option<Weak<dyn Gate>>,
/// A std mutex: never held across an await.
table: Mutex<Table>,
}
impl Paths {
- /// Over `transports`, by scheme, in order of preference; PINGs and PONGs counted in
- /// `traffic`.
+ /// Over `transports`, by scheme, in order of preference; its own PINGs and PONGs
+ /// counted in `traffic`, the firewall through `gate`.
pub(crate) fn new(
transports: Vec<(&'static str, Arc<dyn Transport>)>,
traffic: Option<Arc<Traffic>>,
+ gate: Option<Weak<dyn Gate>>,
) -> Self {
Self {
links: transports
@@ -317,10 +345,15 @@ impl Paths {
.map(|(scheme, inner)| Link { scheme, inner })
.collect(),
traffic,
+ gate,
table: Mutex::default(),
}
}
+ fn gate(&self) -> Option<Arc<dyn Gate>> {
+ self.gate.as_ref()?.upgrade()
+ }
+
/// The transport `address` is on, and the transport identity at it
/// (`Transport::locate`, which notes where that is if the address says); None if no
/// transport of ours has its scheme, or it isn't one of that transport's.
@@ -519,6 +552,9 @@ impl Paths {
Ok(()) => {
let mut table = self.table.lock_or_recover();
table.sent(address);
+ if let Some(ParsedMessage::Ping(nonce)) = message.message.as_deref() {
+ table.pinged(address, *nonce, Instant::now());
+ }
if !failures.is_empty() {
table.counts.failovers += 1;
drop(table);
@@ -545,8 +581,15 @@ impl Paths {
}
}
- /// Sends `message` to its `address`.
+ /// Sends `message` to its `address` (dropped silently if the firewall blocks it).
pub(crate) async fn send(&self, message: &EncodedMessage) -> Result<()> {
+ if self
+ .gate()
+ .is_some_and(|gate| gate.blocks_outbound(&message.address))
+ {
+ tracing::trace!(to = %message.address, "Firewall dropped outbound message");
+ return Ok(());
+ }
let (link, target) = self
.split(&message.address)
.with_context(|| format!("no transport for {}", message.address))?;
@@ -583,7 +626,8 @@ impl Paths {
stats
}
- /// Sends a PING or PONG with `nonce` to `address`, counted (`traffic`).
+ /// Sends a PING (or with `pong`, a PONG) with `nonce` to `address`, counted
+ /// (`traffic`).
async fn send_probe(&self, address: &str, pong: bool, nonce: u64) -> Result<()> {
let message = if pong {
ParsedMessage::Pong(nonce)
@@ -622,8 +666,8 @@ impl Paths {
}
}
- /// A message from `address`: a PING is answered, a PONG noted; anything else goes to
- /// `handler`.
+ /// A message from `address`, handed on to `handler`: a PING answered and a PONG
+ /// noted first; dropped if the firewall blocks it.
async fn receive(
self: Arc<Self>,
address: String,
@@ -631,29 +675,31 @@ impl Paths {
payload_len: usize,
handler: &PathHandler,
) -> Result<()> {
+ if self
+ .gate()
+ .is_some_and(|gate| gate.blocks_inbound(&address))
+ {
+ tracing::trace!(from = %address, "Firewall dropped inbound message");
+ return Ok(());
+ }
match message {
ParsedMessage::Ping(nonce) => {
- if let Some(traffic) = &self.traffic {
- traffic.received(&address, MessageKind::Ping, payload_len);
- }
+ let paths = Arc::clone(&self);
+ let to = address.clone();
tokio::spawn(async move {
- if let Err(error) = self.send_probe(&address, true, nonce).await {
- debug!(to = %address, error = %format!("{error:#}"), "Couldn't answer a PING");
+ if let Err(error) = paths.send_probe(&to, true, nonce).await {
+ debug!(to = %to, error = %format!("{error:#}"), "Couldn't answer a PING");
}
});
- Ok(())
}
ParsedMessage::Pong(nonce) => {
- if let Some(traffic) = &self.traffic {
- traffic.received(&address, MessageKind::Pong, payload_len);
- }
self.table
.lock_or_recover()
.pong(&address, nonce, Instant::now());
- Ok(())
}
- message => handler(address, message, payload_len).await,
+ _ => {}
}
+ handler(address, message, payload_len).await
}
/// Every transport's consumer, each message handed to `handler` under the sender's
@@ -712,7 +758,7 @@ impl Paths {
}
/// A PING's nonce: random, so a PONG can't be made up for a PING not seen.
-fn random_nonce() -> u64 {
+pub(crate) fn random_nonce() -> u64 {
let mut bytes = [0u8; 8];
if getrandom::fill(&mut bytes).is_err() {
// No randomness: a nonce still tells answers apart, if not unguessably.
@@ -764,6 +810,7 @@ mod tests {
("other", second.transport(id) as Arc<dyn Transport>),
],
None,
+ None,
))
};
(on_both("a"), on_both("b"))
@@ -896,6 +943,99 @@ mod tests {
assert!(table.paths.is_empty() && table.chosen.is_empty());
}
+ /// The node's keepalive PING on a path is its probe: paths doesn't ping it too, and
+ /// the PONG measures it.
+ #[test]
+ fn keepalive_pings_are_probes() {
+ let mut table = Table::default();
+ let now = Instant::now();
+ let mut nonce = 100;
+ let mut next = || {
+ nonce += 1;
+ nonce
+ };
+ table.rank(paths(&["libp2p:x", "mycelium:x"]), now);
+ table.pinged("libp2p:x", 7, now);
+ let due = table.due(now, &mut next);
+ assert_eq!(due.len(), 1);
+ assert_eq!(due[0].0, "mycelium:x");
+ assert!(
+ table
+ .pong("libp2p:x", 7, now + Duration::from_millis(20))
+ .is_some()
+ );
+ assert_eq!(table.counts.probes_sent, 2);
+ // A peer with one path isn't measured.
+ table.pinged("libp2p:z", 8, now);
+ assert!(!table.paths.contains_key("libp2p:z"));
+ }
+
+ /// Blocks one address, both ways.
+ struct Blocking(&'static str);
+
+ impl Gate for Blocking {
+ fn blocks_inbound(&self, address: &str) -> bool {
+ address == self.0
+ }
+ fn blocks_outbound(&self, address: &str) -> bool {
+ address == self.0
+ }
+ }
+
+ /// The firewall drops what comes from a blocked address before anything sees it
+ /// (a PING isn't answered), and sends to one silently.
+ #[tokio::test]
+ async fn the_firewall_drops_both_ways() {
+ let net = SimNet::new(4);
+ let gate: Arc<dyn Gate> = Arc::new(Blocking("sim:a"));
+ let b = Arc::new(Paths::new(
+ vec![("sim", net.transport("b") as Arc<dyn Transport>)],
+ None,
+ Some(Arc::downgrade(&gate)),
+ ));
+ let a = Arc::new(Paths::new(
+ vec![("sim", net.transport("a") as Arc<dyn Transport>)],
+ None,
+ None,
+ ));
+ let received = |paths: &Arc<Paths>| {
+ let (sender, received) = tokio::sync::mpsc::unbounded_channel();
+ let handler: Arc<PathHandler> = Arc::new(move |from, message, _| {
+ let sender = sender.clone();
+ Box::pin(async move {
+ let _ = sender.send((from, message.kind()));
+ Ok(())
+ })
+ });
+ let paths = Arc::clone(paths);
+ tokio::spawn(async move { paths.start_consumer(handler).await });
+ received
+ };
+ let (mut at_a, mut at_b) = (received(&a), received(&b));
+
+ // a's PING never reaches b, so nothing answers it.
+ let ping = a.encode("sim:b", &ParsedMessage::Ping(5)).unwrap();
+ a.send(&ping).await.unwrap();
+ // b's send to a "works", and goes nowhere.
+ let status = b.encode("sim:a", &some_status()).unwrap();
+ b.send(&status).await.unwrap();
+ tokio::time::sleep(Duration::from_millis(200)).await;
+ assert!(at_b.try_recv().is_err());
+ assert!(at_a.try_recv().is_err());
+
+ // Without the gate: a PING is answered by b's paths, and goes on up there.
+ drop(gate);
+ a.send(&ping).await.unwrap();
+ assert_eq!(
+ at_b.recv().await.unwrap(),
+ ("sim:a".to_owned(), crate::transport::MessageKind::Ping)
+ );
+ assert_eq!(
+ at_a.recv().await.unwrap(),
+ ("sim:b".to_owned(), crate::transport::MessageKind::Pong)
+ );
+ }
+
/// A transport whose sends all fail.
struct Broken;
@@ -948,10 +1088,12 @@ mod tests {
("sim", net.transport("a") as Arc<dyn Transport>),
],
None,
+ None,
);
let b = Arc::new(Paths::new(
vec![("sim", net.transport("b") as Arc<dyn Transport>)],
None,
+ None,
));
let (sender, mut received) = tokio::sync::mpsc::unbounded_channel();
let handler: Arc<PathHandler> = Arc::new(move |from, _, _| {
diff --git a/src/transport/traffic.rs b/src/transport/traffic.rs
index 9b5f9c9..b070023 100644
--- a/src/transport/traffic.rs
+++ b/src/transport/traffic.rs
@@ -5,8 +5,8 @@
//! (for Mycelium: the topic, or the base64 of its HTTP API). Counters only grow (from
//! the start of the process); rates are for whoever reads them.
//!
-//! A message dropped by the debug firewall still counts as sent (the firewall plays a
-//! lossy network) but not as received.
+//! A message dropped by the debug firewall (in the paths layer, below) still counts as
+//! sent (the firewall plays a lossy network) but not as received.
use std::collections::HashMap;
use std::sync::{Arc, Mutex, MutexGuard};
@@ -188,6 +188,10 @@ impl Transport for CountingTransport {
self.inner.stats()
}
+ fn answers_pings(&self) -> bool {
+ self.inner.answers_pings()
+ }
+
async fn resolve_seed(&self, seed: &str) -> Result<String> {
self.inner.resolve_seed(seed).await
}
generated by cgit v1.3.1 (git 2.54.0) at 2026-10-28 18:45:51 +0000