Runtime diagnostics for leaked messages in unbounded channels (part 2) (

paritytech#13020)

* Fix code review issues

* Clarify doc

* Get rid of backtrace mutex

* kick CI

client/network/Cargo.toml

@@ -17,6 +17,7 @@ targets = ["x86_64-unknown-linux-gnu"]
 array-bytes = "4.1"
 async-trait = "0.1"
 asynchronous-codec = "0.6"
+backtrace = "0.3.67"
 bytes = "1"
 codec = { package = "parity-scale-codec", version = "3.0.0", features = ["derive"] }
 either = "1.5.3"

client/network/src/service/out_events.rs

@@ -31,13 +31,13 @@
 //! - Send events by calling [`OutChannels::send`]. Events are cloned for each sender in the
 //! collection.
 
+use backtrace::Backtrace;
 use futures::{channel::mpsc, prelude::*, ready, stream::FusedStream};
 use log::error;
 use parking_lot::Mutex;
 use prometheus_endpoint::{register, CounterVec, GaugeVec, Opts, PrometheusError, Registry, U64};
 use sc_network_common::protocol::event::Event;
 use std::{
-	backtrace::{Backtrace, BacktraceStatus},
 	cell::RefCell,
 	fmt,
 	pin::Pin,
@@ -62,7 +62,7 @@ pub fn channel(name: &'static str, queue_size_warning: i64) -> (Sender, Receiver
 		queue_size: queue_size.clone(),
 		queue_size_warning,
 		warning_fired: false,
-		creation_backtrace: Backtrace::capture(),
+		creation_backtrace: Backtrace::new_unresolved(),
 		metrics: metrics.clone(),
 	};
 	let rx = Receiver { inner: rx, name, queue_size, metrics };
@@ -91,7 +91,8 @@ pub struct Sender {
 	warning_fired: bool,
 	/// Backtrace of a place where the channel was created.
 	creation_backtrace: Backtrace,
-	/// Clone of [`Receiver::metrics`].
+	/// Clone of [`Receiver::metrics`]. Will be initialized when [`Sender`] is added to
+	/// [`OutChannels`] with `OutChannels::push()`.
 	metrics: Arc<Mutex<Option<Arc<Option<Metrics>>>>>,
 }
 
@@ -193,17 +194,12 @@ impl OutChannels {
 			let queue_size = sender.queue_size.fetch_add(1, Ordering::Relaxed);
 			if queue_size == sender.queue_size_warning && !sender.warning_fired {
 				sender.warning_fired = true;
-				match sender.creation_backtrace.status() {
-					BacktraceStatus::Captured => error!(
-						"The number of unprocessed events in channel `{}` reached {}.\n\
-						 The channel was created at:\n{}",
-						sender.name, sender.queue_size_warning, sender.creation_backtrace,
-					),
-					_ => error!(
-						"The number of unprocessed events in channel `{}` reached {}.",
-						sender.name, sender.queue_size_warning,
-					),
-				}
+				sender.creation_backtrace.resolve();
+				error!(
+					"The number of unprocessed events in channel `{}` reached {}.\n\
+					 The channel was created at:\n{:?}",
+					sender.name, sender.queue_size_warning, sender.creation_backtrace,
+				);
 			}
 			sender.inner.unbounded_send(event.clone()).is_ok()
 		});

client/utils/Cargo.toml

@@ -10,6 +10,7 @@ description = "I/O for Substrate runtimes"
 readme = "README.md"
 
 [dependencies]
+backtrace = "0.3.67"
 futures = "0.3.21"
 futures-timer = "3.0.2"
 lazy_static = "1.4.0"

client/utils/src/mpsc.rs

@@ -37,6 +37,7 @@ mod inner {
 mod inner {
 	// tracing implementation
 	use crate::metrics::UNBOUNDED_CHANNELS_COUNTER;
+	use backtrace::Backtrace;
 	use futures::{
 		channel::mpsc::{
 			self, SendError, TryRecvError, TrySendError, UnboundedReceiver, UnboundedSender,
@@ -47,7 +48,6 @@ mod inner {
 	};
 	use log::error;
 	use std::{
-		backtrace::{Backtrace, BacktraceStatus},
 		pin::Pin,
 		sync::{
 			atomic::{AtomicBool, AtomicI64, Ordering},
@@ -108,7 +108,7 @@ mod inner {
 			queue_size: queue_size.clone(),
 			queue_size_warning,
 			warning_fired: Arc::new(AtomicBool::new(false)),
-			creation_backtrace: Arc::new(Backtrace::capture()),
+			creation_backtrace: Arc::new(Backtrace::new_unresolved()),
 		};
 		let receiver = TracingUnboundedReceiver { inner: r, name, queue_size };
 		(sender, receiver)
@@ -149,23 +149,20 @@ mod inner {
 
 				let queue_size = self.queue_size.fetch_add(1, Ordering::Relaxed);
 				if queue_size == self.queue_size_warning &&
-					!self.warning_fired.load(Ordering::Relaxed)
+					self.warning_fired
+						.compare_exchange(false, true, Ordering::Relaxed, Ordering::Relaxed)
+						.is_ok()
 				{
 					// `warning_fired` and `queue_size` are not synchronized, so it's possible
 					// that the warning is fired few times before the `warning_fired` is seen
 					// by all threads. This seems better than introducing a mutex guarding them.
-					self.warning_fired.store(true, Ordering::Relaxed);
-					match self.creation_backtrace.status() {
-						BacktraceStatus::Captured => error!(
-							"The number of unprocessed messages in channel `{}` reached {}.\n\
-							 The channel was created at:\n{}",
-							self.name, self.queue_size_warning, self.creation_backtrace,
-						),
-						_ => error!(
-							"The number of unprocessed messages in channel `{}` reached {}.",
-							self.name, self.queue_size_warning,
-						),
-					}
+					let mut backtrace = (*self.creation_backtrace).clone();
+					backtrace.resolve();
+					error!(
+						"The number of unprocessed messages in channel `{}` reached {}.\n\
+						 The channel was created at:\n{:?}",
+						self.name, self.queue_size_warning, backtrace,
+					);
 				}
 
 				s