vm: "afterEvaluate", evaluate() return a promise from the outer context

Consider the default context A with a microtask queue QA, and a context B with its own microtask queue QB. Context B is constructed with vm.createContext(..., {microtaskMode: "afterEvaluate"}). The evaluation in context B can be performed via vm.Script or vm.SourceTextModule. The standard (https://tc39.es/ecma262/#sec-newpromiseresolvethenablejob) dictates that, when resolving a {promise} with {resolution}, from any context, the {then} method on {promise} should be called within a task enqueued on the microtask queue from the context associated with {then}. Specifically, after evaluating a script or module in context B, any promises created within B, if later resolved within A, will result in a task to be enqueued back onto QB, even long after we are done evaluating any code within B. This creates a challenge for users of node:vm in "afterEvaluate" mode. In ContextifyScript::EvalMachine() and in ModuleWrap::Evaluate(), we only drain the microtask queue QB a single time after running the script or evaluating the module. After that point, the queue will not be drained unless another script or module is evaluated in the same context. In the following scenario, prior to this patch, the log statement will not be printed:	const microtaskMode = "afterEvaluate";	const context = vm.createContext({}, {microtaskMode});	const source = "";	const module = new vm.SourceTextModule(source, {context});	await module.link(() => null);	await module.evaluate();	console.log("NOT PRINTED"); Within `evaluate()`, there is this `await` statement:	await this[kWrap].evaluate(timeout, breakOnSigint) Since the promise returned by ModuleWrap::Evaluate() is the top-level capability for {module}, a promise created within B, V8 will enqueue a task on QB. But since this is after the PerformCheckpoint() call in ModuleWrap::Evaluate(), the task in QB is never run. In the meantime, since QA is empty, the Node process simply exits (with a warning about the unsettled promise, if it happened to be a top-level await). While being unable to do `await module.evaluate()` is clearly a problem, more generally, it is intended that in "afterEvaluate" mode, promises created in the inner context cannot make progress if, and until, the microtask queue of the inner context is checkpointed. Therefore, to address this issue, the fix is narrow: When the module has its own microtask queue, i.e. in "afterEvaluate" mode, the inner-context promise returned by v8::SourceTextModule::Evaluate() is first resolved to an outer-context promise, then we checkpoint the microtask queue of the inner context, then we return the outer-context promise we just built. This ensures that in the statement `await this[kWrap].evaluate(...)`, the promise returned can be resolved within the outer context, without involving the microtask queue in the inner context. Fixes: #59541 Refs: https://issues.chromium.org/issues/441679231 Refs: https://groups.google.com/g/v8-dev/c/YIeRg8CUNS8/m/rEQdFuNZAAAJ PR-URL: #59801 Refs: https://tc39.es/ecma262/#sec-newpromiseresolvethenablejob Reviewed-By: Anna Henningsen <anna@addaleax.net>

Author: ericrannaud

src/module_wrap.cc

@@ -759,8 +759,48 @@ void ModuleWrap::Evaluate(const FunctionCallbackInfo<Value>& args) {
  MaybeLocal<Value> result;
  auto run = [&]() {
  MaybeLocal<Value> result = module->Evaluate(context);
- if (!result.IsEmpty() && microtask_queue)
+
+ Local<Value> res;
+ if (result.ToLocal(&res) && microtask_queue) {
+ DCHECK(res->IsPromise());
+
+ // To address https://github.com/nodejs/node/issues/59541 when the
+ // module has its own separate microtask queue in microtaskMode
+ // "afterEvaluate", we avoid returning a promise built inside the
+ // module's own context.
+ //
+ // Instead, we build a promise in the outer context, which we resolve
+ // with {result}, then we checkpoint the module's own queue, and finally
+ // we return the outer-context promise.
+ //
+ // If we simply returned the inner promise {result} directly, per
+ // https://tc39.es/ecma262/#sec-newpromiseresolvethenablejob, the outer
+ // context, when resolving a promise coming from a different context,
+ // would need to enqueue a task (known as a thenable job task) onto the
+ // queue of that different context (the module's context). But this queue
+ // will normally not be checkpointed after evaluate() returns.
+ //
+ // This means that the execution flow in the outer context would
+ // silently fall through at the statement (in lib/internal/vm/module.js):
+ // await this[kWrap].evaluate(timeout, breakOnSigint)
+ //
+ // This is true for any promises created inside the module's context
+ // and made available to the outer context, as the node:vm doc explains.
+ //
+ // We must handle this particular return value differently to make it
+ // possible to await on the result of evaluate().
+ Local<Context> outer_context = isolate->GetCurrentContext();
+ Local<Promise::Resolver> resolver;
+ if (!Promise::Resolver::New(outer_context).ToLocal(&resolver)) {
+ return MaybeLocal<Value>();
+ }
+ if (resolver->Resolve(outer_context, res).IsNothing()) {
+ return MaybeLocal<Value>();
+ }
+ result = resolver->GetPromise();
+
  microtask_queue->PerformCheckpoint(isolate);
+ }
  return result;
  };
  if (break_on_sigint && timeout != -1) {

test/parallel/test-vm-module-after-evaluate.js

@@ -14,7 +14,7 @@ const microtaskMode = 'afterEvaluate';
 
 (async () => {
  const mustNotCall1 = common.mustNotCall();
- const mustNotCall2 = common.mustNotCall();
+ const mustCall1 = common.mustCall();
 
  const inner = {};
 
@@ -28,7 +28,6 @@ const microtaskMode = 'afterEvaluate';
  await module.link(mustNotCall1);
  await module.evaluate();
 
- // This is Issue 59541, the next statement is not executed, of course
- // it should be.
- mustNotCall2();
-})().then(common.mustNotCall());
+ // Prior to the fix for Issue 59541, the next statement was never executed.
+ mustCall1();
+})().then(common.mustCall());