Wait for overlapped WriteFile completion in Rust (RustPython#7383)

youknowone · youknowone · commit ca69664b104b · 2026-03-08T18:28:33.000+09:00
On Windows, SimpleQueue skips write locking because pipe
writes are assumed atomic. Without GIL, PipeConnection.
_send_bytes races on _send_ov when multiple threads call
send_bytes concurrently (e.g. _terminate_pool vs workers).

Wait for pending overlapped writes inside WriteFile before
returning to Python, so ERROR_IO_PENDING is never exposed
and the _send_ov assignment window is negligible.
diff --git a/crates/vm/src/stdlib/_winapi.rs b/crates/vm/src/stdlib/_winapi.rs
@@ -1251,6 +1251,26 @@ mod _winapi {
 
                 err
             };
+
+            // Without GIL, the Python-level PipeConnection._send_bytes has a
+            // race on _send_ov when the caller (SimpleQueue) skips locking on
+            // Windows. Wait for completion here so the caller never sees
+            // ERROR_IO_PENDING and never blocks in WaitForMultipleObjects,
+            // keeping the _send_ov window negligibly small.
+            if err == ERROR_IO_PENDING {
+                let event = ov.inner.lock().overlapped.hEvent;
+                vm.allow_threads(|| unsafe {
+                    windows_sys::Win32::System::Threading::WaitForSingleObject(
+                        event,
+                        windows_sys::Win32::System::Threading::INFINITE,
+                    );
+                });
+                let result = vm
+                    .ctx
+                    .new_tuple(vec![ov.into_pyobject(vm), vm.ctx.new_int(0u32).into()]);
+                return Ok(result.into());
+            }
+
             let result = vm
                 .ctx
                 .new_tuple(vec![ov.into_pyobject(vm), vm.ctx.new_int(err).into()]);
