fs: clean up Dir.read() uv_fs_t data before calling into JS

A call into JS can schedule another operation on the same `uv_dir_t`.
In particular, when the handle is closed from the callback for a
directory read operation, there previously was a race condition window:

1. A `dir.read()` operation is submitted to libuv
2. The read operation is finished by libuv, calling `AfterDirRead()`
3. We call into JS
4. JS calls dir.close()
5. libuv posts the close request to a thread in the pool
6. The close request runs, destroying the directory handle
7. `AfterDirRead()` is being exited.

Exiting the `FSReqAfterScope` in step 7 attempts to destroy the original
uv_fs_t` from step 1, which now points to an `uv_dir_t` that has
already been destroyed in step 5.

By forcing the `FSReqAfterScope` to clean up before we call into JS,
we can be sure that no other operations on the same `uv_dir_t` are
submitted concurrently.

This addresses issues observed when running with ASAN/valgrind.

PR-URL: #33274
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>

Author: addaleax

src/node_dir.cc

@@ -197,8 +197,8 @@ static MaybeLocal<Array> DirentListToArray(
 }
 
 static void AfterDirRead(uv_fs_t* req) {
-  FSReqBase* req_wrap = FSReqBase::from_req(req);
-  FSReqAfterScope after(req_wrap, req);
+  BaseObjectPtr<FSReqBase> req_wrap { FSReqBase::from_req(req) };
+  FSReqAfterScope after(req_wrap.get(), req);
 
   if (!after.Proceed()) {
     return;
@@ -210,12 +210,12 @@ static void AfterDirRead(uv_fs_t* req) {
   if (req->result == 0) {
     // Done
     Local<Value> done = Null(isolate);
+    after.Clear();
     req_wrap->Resolve(done);
     return;
   }
 
   uv_dir_t* dir = static_cast<uv_dir_t*>(req->ptr);
-  req->ptr = nullptr;
 
   Local<Value> error;
   Local<Array> js_array;
@@ -224,9 +224,13 @@ static void AfterDirRead(uv_fs_t* req) {
                          req->result,
                          req_wrap->encoding(),
                          &error).ToLocal(&js_array)) {
+    // Clear libuv resources *before* delivering results to JS land because
+    // that can schedule another operation on the same uv_dir_t. Ditto below.
+    after.Clear();
     return req_wrap->Reject(error);
   }
 
+  after.Clear();
   req_wrap->Resolve(js_array);
 }
 

src/node_file.cc

@@ -540,8 +540,15 @@ FSReqAfterScope::FSReqAfterScope(FSReqBase* wrap, uv_fs_t* req)
 }
 
 FSReqAfterScope::~FSReqAfterScope() {
+  Clear();
+}
+
+void FSReqAfterScope::Clear() {
+  if (!wrap_) return;
+
   uv_fs_req_cleanup(wrap_->req());
-  delete wrap_;
+  wrap_->Detach();
+  wrap_.reset();
 }
 
 // TODO(joyeecheung): create a normal context object, and
@@ -554,12 +561,16 @@ FSReqAfterScope::~FSReqAfterScope() {
 // which is also why the errors should have been constructed
 // in JS for more flexibility.
 void FSReqAfterScope::Reject(uv_fs_t* req) {
-  wrap_->Reject(UVException(wrap_->env()->isolate(),
-                            req->result,
-                            wrap_->syscall(),
-                            nullptr,
-                            req->path,
-                            wrap_->data()));
+  BaseObjectPtr<FSReqBase> wrap { wrap_ };
+  Local<Value> exception =
+      UVException(wrap_->env()->isolate(),
+                  req->result,
+                  wrap_->syscall(),
+                  nullptr,
+                  req->path,
+                  wrap_->data());
+  Clear();
+  wrap->Reject(exception);
 }
 
 bool FSReqAfterScope::Proceed() {

src/node_file.h

@@ -157,6 +157,7 @@ class FSReqAfterScope final {
  public:
   FSReqAfterScope(FSReqBase* wrap, uv_fs_t* req);
   ~FSReqAfterScope();
+  void Clear();
 
   bool Proceed();
 
@@ -168,7 +169,7 @@ class FSReqAfterScope final {
   FSReqAfterScope& operator=(const FSReqAfterScope&&) = delete;
 
  private:
-  FSReqBase* wrap_ = nullptr;
+  BaseObjectPtr<FSReqBase> wrap_;
   uv_fs_t* req_ = nullptr;
   v8::HandleScope handle_scope_;
   v8::Context::Scope context_scope_;