fix: start new connections independently of connection failures

With current implementation
every time connection fails
we take the next delay from `delays` iterator.
In the worst case first 4 DNS results
immediately refuse connection
and we start fifth connection attempt
with 1 year timeout,
effectively continuing all remaining
connection attempts without concurrency.

With new implementation
new connection attempts are
added to `connection_attempt_set`
independently of connection failures
and after 10 seconds
we always end up with five
parallel connection attempts
as long as there are enough IP addresses.

src/net.rs

@@ -152,13 +152,15 @@ where
 
     // Start additional connection attempts after 300 ms, 1 s, 5 s and 10 s.
     // This way we can have up to 5 parallel connection attempts at the same time.
-    let mut delays = [
+    let mut delay_set = JoinSet::new();
+    for delay in [
         Duration::from_millis(300),
         Duration::from_secs(1),
         Duration::from_secs(5),
         Duration::from_secs(10),
-    ]
+    ] {
-    .into_iter();
+        delay_set.spawn(tokio::time::sleep(delay));
+    }
 
     let mut first_error = None;
 
@@ -167,14 +169,10 @@ where
             connection_attempt_set.spawn(fut);
         }
 
-        let one_year = Duration::from_secs(60 * 60 * 24 * 365);
+        tokio::select! {
-        let delay = delays.next().unwrap_or(one_year); // one year can be treated as infinitely long here
+            biased;
-        let Ok(res) = timeout(delay, connection_attempt_set.join_next()).await else {
-            // The delay for starting the next connection attempt has expired.
-            // `continue` the loop to push the next connection into connection_attempt_set.
-            continue;
-        };
 
+            res = connection_attempt_set.join_next() => {
                 match res {
                     Some(res) => {
                         match res.context("Failed to join task") {
@@ -200,11 +198,25 @@ where
                         );
                     }
                 }
+            },
+
+            _ = delay_set.join_next(), if !delay_set.is_empty() => {
+                // Delay expired.
+                //
+                // Don't do anything other than pushing
+                // another connection attempt into `connection_attempt_set`.
+            }
+        }
     };
 
     // Abort remaining connection attempts and free resources
     // such as OS sockets and `Context` references
     // held by connection attempt tasks.
+    //
+    // `delay_set` contains just `sleep` tasks
+    // so no need to await futures there,
+    // it is enough that futures are aborted
+    // when the set is dropped.
     connection_attempt_set.shutdown().await;
 
     res