fix: make WasmBase::doAfterVmCallActions reentry-safe (drain-to-local)

include/proxy-wasm/wasm.h

@@ -147,9 +147,21 @@ class WasmBase : public std::enable_shared_from_this<WasmBase> {
     // NB: this may be deleted by a delayed function unless prevented.
     if (!after_vm_call_actions_.empty()) {
       auto self = shared_from_this();
-      while (!self->after_vm_call_actions_.empty()) {
-        auto f = std::move(self->after_vm_call_actions_.front());
-        self->after_vm_call_actions_.pop_front();
+      // Swap the queue into a local before draining, then iterate the snapshot exactly
+      // once. Actions re-added by a callback during the drain land in the (now-empty)
+      // member queue and are picked up by the next-outer DeferAfterCallActions frame,
+      // instead of being re-run in this same loop. With the old unguarded
+      // `while (!empty())`, an action that (directly or transitively) re-enqueues during
+      // the drain -- e.g. a host call such as sendLocalReply/injectEncodedDataToFilterChain
+      // that synchronously re-enters the VM and schedules another after-VM-call action --
+      // means the loop never observes an empty queue and spins forever, pinning a CPU at
+      // 100%. Deferring re-added actions to the next-outer frame also prevents an action
+      // queued in one phase (e.g. continueStream from onRequestHeaders) from executing
+      // inside a nested VM call of a different phase (e.g. onResponseHeaders triggered by
+      // sendLocalReply); see proxy-wasm/proxy-wasm-cpp-host#326.
+      std::deque<std::function<void()>> local;
+      local.swap(self->after_vm_call_actions_);
+      for (auto &f : local) {
         f();
       }
     }

test/wasm_test.cc

@@ -324,4 +324,32 @@ TEST_P(TestVm, CleanupThreadLocalCacheKeys) {
   EXPECT_TRUE(stale_wasms_keys.empty());
 }
 
+// doAfterVmCallActions() must not spin forever when a queued action re-enqueues itself
+// during the drain. With the old unguarded `while (!empty())` loop this loops forever;
+// with the drain-to-local fix each call runs the snapshot exactly once and the re-added
+// copy is deferred to the next drain.
+TEST_P(TestVm, DoAfterVmCallActionsReentrySafe) {
+  // WasmBase::doAfterVmCallActions() calls shared_from_this(), so the instance must be
+  // owned by a std::shared_ptr (a stack WasmBase would crash).
+  auto wasm = std::make_shared<TestWasm>(makeVm(engine_));
+
+  int count = 0;
+  std::function<void()> f;
+  f = [&]() {
+    ++count;
+    // Re-enqueue self during the drain. Under the old code this would be picked up by the
+    // same loop and never terminate.
+    wasm->addAfterVmCallAction(f);
+  };
+
+  wasm->addAfterVmCallAction(f);
+  // Snapshot {f} runs exactly once; the re-added copy lands in the now-empty member queue.
+  wasm->doAfterVmCallActions();
+  EXPECT_EQ(count, 1); // Returns after one pass (the old while-loop would never return).
+
+  // The re-added copy runs on the next frame (and re-enqueues once more).
+  wasm->doAfterVmCallActions();
+  EXPECT_EQ(count, 2);
+}
+
 } // namespace proxy_wasm