fix(focas): a caller's own cancellation during connect does not open the backoff window; deterministic concurrent-connect test

Review follow-up on c940349f: a cancelled caller recorded a connect failure, throttling every caller queued
on ConnectGate for ~1 s. The 200 ms settle delay in the concurrent-connect test is replaced by a
TaskCompletionSource the fake connect signals on entry. Both tests verified red with their fix reverted.

Claude-Session: https://claude.ai/code/session_01Ef2ozxS7QPanGT8uoATbTe
This commit is contained in:
Joseph Doherty
2026-10-08 21:28:00 -04:00
parent c940349fe8
commit 54a00fd760
2 changed files with 49 additions and 6 deletions
@@ -1855,12 +1855,15 @@ public sealed class FocasDriver : IDriver, IReadable, IWritable, ITagDiscovery,
{
await client.ConnectAsync(device.ParsedAddress, _options.Timeout, ct).ConfigureAwait(false);
}
catch
catch (Exception ex)
{
// Dispose only OUR client — never one another path installed meanwhile (recycle/teardown).
client.Dispose();
if (ReferenceEquals(device.Client, client)) device.Client = null;
device.Backoff.RecordFailure(DateTime.UtcNow); // 05/STAB-8: open the backoff window
// The caller's OWN cancellation says nothing about the CNC: do not open the backoff window, or
// every caller queued on ConnectGate behind it would be throttled for ~1 s for nothing.
if (!(ex is OperationCanceledException && ct.IsCancellationRequested))
device.Backoff.RecordFailure(DateTime.UtcNow); // 05/STAB-8: open the backoff window
throw;
}
device.Backoff.RecordSuccess(); // 05/STAB-8: recovery resets the window immediately
@@ -19,12 +19,13 @@ public sealed class FocasConcurrentConnectTests
{
/// <summary>A fake whose connect parks until released, and which — like the real wire client — refuses to
/// finish connecting once disposed.</summary>
private sealed class GatedConnectClient(Task release) : FakeFocasClient
private sealed class GatedConnectClient(Task release, TaskCompletionSource? entered = null) : FakeFocasClient
{
private bool _disposed;
public override async Task ConnectAsync(FocasHostAddress address, TimeSpan timeout, CancellationToken ct)
{
entered?.TrySetResult();
await release.WaitAsync(ct);
if (_disposed) throw new ObjectDisposedException("FocasWireClient");
await base.ConnectAsync(address, timeout, ct);
@@ -42,7 +43,8 @@ public sealed class FocasConcurrentConnectTests
public async Task Concurrent_first_reads_open_one_session_and_both_succeed()
{
var release = new TaskCompletionSource(TaskCreationOptions.RunContinuationsAsynchronously);
var factory = new FakeFocasClientFactory { Customise = () => new GatedConnectClient(release.Task) };
var entered = new TaskCompletionSource(TaskCreationOptions.RunContinuationsAsynchronously);
var factory = new FakeFocasClientFactory { Customise = () => new GatedConnectClient(release.Task, entered) };
await using var drv = new FocasDriver(new FocasDriverOptions
{
Devices = [new FocasDeviceOptions("focas://10.0.0.5:8193")],
@@ -55,9 +57,11 @@ public sealed class FocasConcurrentConnectTests
await drv.InitializeAsync("{}", CancellationToken.None);
var first = drv.ReadAsync(["A"], CancellationToken.None);
// Deterministic: the first caller is now parked INSIDE its connect (client created, not yet connected) —
// the exact window the race needs. The second caller then runs synchronously up to its first real await:
// pre-fix that is "dispose the first client, Create() a second"; with the gate it parks on ConnectGate.
await entered.Task.WaitAsync(TimeSpan.FromSeconds(5));
var second = drv.ReadAsync(["B"], CancellationToken.None);
// Give both callers time to reach the connect path before the (single) connect is allowed to finish.
await Task.Delay(200);
release.SetResult();
var results = await Task.WhenAll(first, second);
@@ -89,4 +93,40 @@ public sealed class FocasConcurrentConnectTests
var read = drv.ReadAsync(["A"], CancellationToken.None);
(await read.WaitAsync(TimeSpan.FromSeconds(5))).Single().StatusCode.ShouldBe(FocasStatusMapper.Good);
}
/// <summary>
/// A caller cancelled during its connect must not open the backoff window: the cancellation is the caller's,
/// not the CNC's, and a recorded failure would throttle every caller queued behind it on the gate.
/// </summary>
[Fact]
public async Task A_caller_cancelled_during_connect_does_not_throttle_the_next_caller()
{
var never = new TaskCompletionSource(TaskCreationOptions.RunContinuationsAsynchronously);
var entered = new TaskCompletionSource(TaskCreationOptions.RunContinuationsAsynchronously);
var factory = new FakeFocasClientFactory();
var created = 0;
// First client parks in connect until cancelled; every later client connects immediately.
factory.Customise = () => Interlocked.Increment(ref created) == 1
? new GatedConnectClient(never.Task, entered)
: new FakeFocasClient();
await using var drv = new FocasDriver(new FocasDriverOptions
{
Devices = [new FocasDeviceOptions("focas://10.0.0.5:8193")],
RawTags = FocasRawTags.Entries([new FocasTagDefinition("A", "focas://10.0.0.5:8193", "R100", FocasDataType.Byte)]),
Probe = new FocasProbeOptions { Enabled = false },
Timeout = TimeSpan.FromSeconds(5),
}, "drv-1", factory);
await drv.InitializeAsync("{}", CancellationToken.None);
using var cts = new CancellationTokenSource();
var cancelled = drv.ReadAsync(["A"], cts.Token);
await entered.Task.WaitAsync(TimeSpan.FromSeconds(5));
cts.Cancel();
try { await cancelled; } catch (OperationCanceledException) { /* either surfaced or mapped — both fine */ }
// Immediately — well inside the 1 s initial backoff window a recorded failure would have opened.
var next = await drv.ReadAsync(["A"], CancellationToken.None).WaitAsync(TimeSpan.FromSeconds(5));
next.Single().StatusCode.ShouldBe(FocasStatusMapper.Good, "a caller's own cancellation must not throttle the next connect");
factory.Clients.Count.ShouldBe(2);
}
}