fix(worker): observe detached drain faults per NEXT-04 discipline

Both tasks the detached-lock-wait path starts and discards now carry the
file's fault-observing continuation, factored out of ObserveAbandonedFault as
ObserveFault so the idiom has one definition. DrainDetachedAsync swallows the
drain but its _writeLock.Release() sits in a finally outside that catch, so a
Release that ever throws — a SemaphoreFullException from some future
double-release regression — had no awaiter and would have surfaced on net48 as
TaskScheduler.UnobservedTaskException at finalization instead of an
attributable failure. Same for a throw out of OnDetachedLockWaitSettled.

Review's Minor (distinguishing a cancelled from a faulted lock wait before
tombstoning) is deliberately not taken: a faulted WaitAsync is unreachable
here — nothing disposes _writeLock — so the branch would be untestable new
logic whose only effect is internal state, the caller already receiving the
fault itself from the rethrow. Recorded as a comment at the site instead.

edited on macOS, windev verification pending (plan Task 11). Re-ran the net10
scratch harness over WorkerFrameWriter and the writer suite: 0 warnings,
31/31 pass.
This commit is contained in:
Joseph Doherty
2026-08-17 04:07:44 -04:00
parent ce5d8ae7c2
commit b8b7b69ba0
@@ -172,6 +172,12 @@ public sealed class WorkerFrameWriter
// caller still observes cancellation while the frame reaches the wire (documented
// above). Rethrow from the wait rather than from the completion, so a claimed frame's
// canceller is not held behind the very write it is abandoning.
//
// The wait ending without the lock IS cancellation in every reachable case — nothing
// disposes _writeLock — so the tombstone's TrySetCanceled is honest. A hypothetical
// faulted wait would take this same path and label the frame cancelled instead of
// faulted; that is internal state only, since the await below rethrows the fault itself
// to the caller.
TombstoneIfUnclaimed(frame, cancellationToken);
await lockWait.ConfigureAwait(false);
}
@@ -343,8 +349,23 @@ public sealed class WorkerFrameWriter
/// <param name="frame">Frame whose completion may fault without an awaiter.</param>
private static void ObserveAbandonedFault(PendingFrame frame)
{
_ = frame.Completion.Task.ContinueWith(
task => _ = task.Exception,
ObserveFault(frame.Completion.Task);
}
/// <summary>
/// Attaches the NEXT-04 fault-observing continuation to a task this writer starts and then
/// discards. Every such task must carry one: with no awaiter, a fault would otherwise reach
/// nobody and resurface as <see cref="TaskScheduler.UnobservedTaskException"/> at finalization
/// — a detached, unattributable failure long after the code that caused it. The continuation
/// runs only on the faulted path and only touches <see cref="Task.Exception"/>, so it can
/// never fault itself; it runs inline because an already-faulted task would otherwise pay a
/// scheduling hop to do nothing.
/// </summary>
/// <param name="task">Discarded task whose fault would otherwise go unobserved.</param>
private static void ObserveFault(Task task)
{
_ = task.ContinueWith(
faulted => _ = faulted.Exception,
CancellationToken.None,
TaskContinuationOptions.OnlyOnFaulted | TaskContinuationOptions.ExecuteSynchronously,
TaskScheduler.Default);
@@ -367,11 +388,14 @@ public sealed class WorkerFrameWriter
/// <param name="lockWait">Outstanding write-lock acquisition the caller has walked away from.</param>
private void DetachLockWait(Task lockWait)
{
_ = lockWait.ContinueWith(
// The continuation task is discarded, so it takes the NEXT-04 fault observer: nothing awaits
// it, and a throw out of OnDetachedLockWaitSettled would otherwise be an unobserved-task
// exception raised at finalization rather than an attributable failure here.
ObserveFault(lockWait.ContinueWith(
OnDetachedLockWaitSettled,
CancellationToken.None,
TaskContinuationOptions.None,
TaskScheduler.Default);
TaskScheduler.Default));
}
/// <summary>
@@ -394,7 +418,11 @@ public sealed class WorkerFrameWriter
return;
}
_ = DrainDetachedAsync();
// Discarded, so it takes the NEXT-04 fault observer too. DrainDetachedAsync swallows the drain
// itself, but its release sits in a finally outside that catch: a Release that ever throws (a
// SemaphoreFullException from some future double-release regression, say) must fail somewhere
// attributable rather than at finalization.
ObserveFault(DrainDetachedAsync());
}
/// <summary>