sql(postgres): ref poll_ref only after request is enqueued in do_run

src/sql_jsc/postgres/PostgresSQLQuery.rs

@@ -479,15 +479,6 @@
         };
         let connection: &PostgresSQLConnection = &connection;
 
-        // `KeepAlive::ref_` takes an `EventLoopCtx` (manual vtable in `bun_io`), not a
-        // `*mut VirtualMachine`. `global_object.bun_vm()` and `get_vm_ctx(.Js)` both
-        // resolve to the same singleton JS VM, so route through the global hook —
-        // identical to `PostgresSQLConnection::vm_ctx`.
-        connection.poll_ref.with_mut(|r| {
-            r.ref_(bun_io::posix_event_loop::get_vm_ctx(
-                bun_io::AllocatorType::Js,
-            ))
-        });
         let query = arguments[1];
 
         if !query.is_object() {
@@ -561,6 +552,16 @@
                 return Err(global_object.throw_out_of_memory());
             }
 
+            // Request is enqueued: keep the event loop alive until the server
+            // responds. KeepAlive is a flag (not a count), so taking this any
+            // earlier would leave it stuck Active on the synchronous-error
+            // returns above.
+            connection.poll_ref.with_mut(|r| {
+                r.ref_(bun_io::posix_event_loop::get_vm_ctx(
+                    bun_io::AllocatorType::Js,
+                ))
+            });
+
             this.this_value.with_mut(|r| r.upgrade(global_object));
             js::target_set_cached(this_value, global_object, query);
             if this.status.get() == Status::Running {
@@ -853,9 +854,19 @@
             .requests
             .with_mut(|q| q.write_item(this_ptr))
             .is_err()
         {
             return Err(global_object.throw_out_of_memory());
         }
+
+        // Request is enqueued: keep the event loop alive until the server
+        // responds. See the matching call in the simple-query branch above
+        // for why this must come after every fallible step.
+        connection.poll_ref.with_mut(|r| {
+            r.ref_(bun_io::posix_event_loop::get_vm_ctx(
+                bun_io::AllocatorType::Js,
+            ))
+        });
+
         this.this_value.with_mut(|r| r.upgrade(global_object));
 
         js::target_set_cached(this_value, global_object, query);

test/js/sql/sql-postgres-run-error-pollref.test.ts

@@ -0,0 +1,31 @@
+// PostgresSQLQuery.do_run refs the connection's poll_ref KeepAlive. KeepAlive
+// is a two-state flag, not a counter, so when this query is the only in-flight
+// work the call flips Inactive -> Active. When do_run then returns early with
+// a synchronous error (bad binding, signature-generation failure, OOM during
+// enqueue, ...) the poll_ref must not be left Active: nothing else on the
+// connection will touch it until the next server message, so the event loop
+// stays pinned and the process never exits.
+//
+// The fixture connects to a mock server, lets the connection go idle, then
+// issues a query whose binding is rejected synchronously before anything is
+// written. It must print the rejection and exit on its own.
+
+import { expect, test } from "bun:test";
+import { bunEnv, bunExe } from "harness";
+import path from "node:path";
+
+test("postgres: synchronous do_run failure does not pin the event loop", async () => {
+  await using proc = Bun.spawn({
+    cmd: [bunExe(), path.join(import.meta.dir, "sql-postgres-run-error-pollref-fixture.ts")],
+    env: bunEnv,
+    stdout: "pipe",
+    stderr: "pipe",
+  });
+
+  const [stdout, stderr, exitCode] = await Promise.all([proc.stdout.text(), proc.stderr.text(), proc.exited]);
+
+  expect({ stdout, stderr }).toEqual({ stdout: "rejected:ERR_INVALID_ARG_TYPE\n", stderr: "" });
+  // exited on its own, not killed by the runner's timeout
+  expect(proc.signalCode).toBeNull();
+  expect(exitCode).toBe(0);
+});