fix(isotp): wake receivers when the reader fails on an exception the generic-catch filter excludes

src/CanKit.Pro.IsoTp/IsoTpChannel.cs

@@ -677,15 +677,45 @@ private async Task RunReaderAsync()
         {
             return; // expected on Dispose
         }
-        catch (Exception ex)
+        catch (Exception ex) when (ex is not OutOfMemoryException
+            and not StackOverflowException
+            and not AccessViolationException
+            and not AppDomainUnloadedException
+            and not BadImageFormatException
+            and not CannotUnloadAppDomainException
+            and not ThreadAbortException)
         {
             lost = ex;
         }
+        catch (Exception ex) when (ex is OutOfMemoryException
+            or StackOverflowException
+            or AccessViolationException
+            or AppDomainUnloadedException
+            or BadImageFormatException
+            or CannotUnloadAppDomainException
+            or ThreadAbortException)
+        {
+            // The filter above is what keeps the generic catch closed, so these used to leave
+            // this method entirely. This method is the reader task. Nothing else observes that
+            // task except Dispose, and Dispose swallows the wait, so the inbox loss was never
+            // published and a ReceiveAsync already waiting never woke. Publish first, then
+            // rethrow: the post is queued before the task faults, and the task still faults.
+            // A stack overflow the runtime detects itself is not delivered here (the process
+            // ends), and ThreadAbortException is not thrown on this runtime. An
+            // OutOfMemoryException from the pump is delivered, and has to take this path.
+            PublishSubscriptionLoss(ex);
+            throw;
+        }
 
-        // Nothing will ever arrive again. A receiver waiting on the inbox would wait for ever, so
-        // the reason is recorded and the receivers are woken: what is buffered is delivered first,
-        // and then every receiver -- not just one -- gets the failure. Sends already fail on their
-        // own, the bus refuses them. On the actor, like everything else that touches the inbox.
+        PublishSubscriptionLoss(lost);
+    }
+
+    // Nothing will ever arrive again. A receiver waiting on the inbox would wait for ever, so
+    // the reason is recorded and the receivers are woken: what is buffered is delivered first,
+    // and then every receiver -- not just one -- gets the failure. Sends already fail on their
+    // own, the bus refuses them. On the actor, like everything else that touches the inbox.
+    private void PublishSubscriptionLoss(Exception lost)
+    {
         try
         {
             // Under the pump lock: a caller that is pumping the subscription right now -- it took

tests/CanKit.Pro.Tests/TestCases/IsoTp/IsoTpChannelIntegrationTests.cs

@@ -1105,6 +1105,58 @@ public async Task A_Failing_Subscription_Ends_The_Inbox_With_Its_Failure()
         lock (reported) reported.Should().ContainSingle().Which.Message.Should().Be("the demux broke");
     }
 
+    // The reader's generic-catch filter does not handle these. They are still reader failures:
+    // a receive that is already waiting, and one that starts afterwards, both have to observe
+    // the same exception, and it has to be raised once. The subscription throws the instance,
+    // so a passing test is the inbox-loss signal, not a stand-in for it. StackOverflowException
+    // and AccessViolationException are left out: an explicit throw is catchable, but a real
+    // stack overflow aborts the process. ThreadAbortException cannot be constructed here.
+    [Theory]
+    [InlineData(nameof(OutOfMemoryException))]
+    [InlineData(nameof(BadImageFormatException))]
+    [InlineData(nameof(AppDomainUnloadedException))]
+    [InlineData(nameof(CannotUnloadAppDomainException))]
+    public async Task An_Excluded_Reader_Failure_Still_Wakes_Waiting_And_Later_Receives(string kind)
+    {
+        var fault = ExcludedReaderFailure(kind);
+        var service = new StarvedReaderBusService { ReaderFault = fault };
+        using var actor = new ProtocolActor();
+        using var channel = new IsoTpChannel(service, IsoTpEndpoint.Normal(0x123, 0x321), FastOptions(),
+            ownsService: false, actor);
+        var reported = new TaskCompletionSource<Exception>(TaskCreationOptions.RunContinuationsAsynchronously);
+        channel.BackgroundExceptionOccurred += (_, ex) => reported.TrySetResult(ex);
+
+        var first = channel.ReceiveAsync();
+        var second = channel.ReceiveAsync();
+        service.WakeReader();
+
+        Func<Task> firstAct = () => first.WaitAsync(ShortTimeout);
+        (await firstAct.Should().ThrowAsync<Exception>()).Which.Should().BeSameAs(fault);
+        Func<Task> secondAct = () => second.WaitAsync(ShortTimeout);
+        (await secondAct.Should().ThrowAsync<Exception>()).Which.Should().BeSameAs(fault);
+
+        (await reported.Task.WaitAsync(ShortTimeout)).Should().BeSameAs(fault);
+
+        Func<Task> later = () => channel.ReceiveAsync().WaitAsync(ShortTimeout);
+        (await later.Should().ThrowAsync<Exception>()).Which.Should().BeSameAs(fault);
+
+        // Publishing the loss must not swallow the failure: the reader task still faults
+        // with the same instance. WaitAsync observes that fault, or times out if it was swallowed.
+        var reader = (Task)typeof(IsoTpChannel).GetField("_readerTask",
+            BindingFlags.Instance | BindingFlags.NonPublic)!.GetValue(channel)!;
+        Func<Task> readerFaulted = () => reader.WaitAsync(ShortTimeout);
+        (await readerFaulted.Should().ThrowAsync<Exception>()).Which.Should().BeSameAs(fault);
+    }
+
+    private static Exception ExcludedReaderFailure(string kind) => kind switch
+    {
+        nameof(OutOfMemoryException) => new OutOfMemoryException("the reader failed"),
+        nameof(BadImageFormatException) => new BadImageFormatException("the reader failed"),
+        nameof(AppDomainUnloadedException) => new AppDomainUnloadedException("the reader failed"),
+        nameof(CannotUnloadAppDomainException) => new CannotUnloadAppDomainException("the reader failed"),
+        _ => throw new ArgumentOutOfRangeException(nameof(kind), kind, "Not an excluded reader failure under test."),
+    };
+
     // A reassembly under way dies with the subscription: it is withdrawn, so
     // GetReceptionsInProgress does not keep reporting a transfer nobody will complete.
     [Fact]