Conceptually what I'm trying to do is very simple. I have a Readable stream in node, and I'm passing that to a native c++ addon where I want to connect that to an IInputStream.
The native library that I'm using works like many c++ (or Java) streaming interfaces that I've seen. The library provides an IInputStream interface (technically an abstract class), which I inherit from and override the virtual functions. Looks like this:
class JsReadable2InputStream : public IInputStream {
public:
// Constructor takes a js v8 object, makes a stream out of it
JsReadable2InputStream(const v8::Local<v8::Object>& streamObj);
~JsReadable2InputStream();
/**
* Blocking read. Blocks until the requested amount of data has been read. However,
* if the stream reaches its end before the requested amount of bytes has been read
* it returns the number of bytes read thus far.
*
* @param begin memory into which read data is copied
* @param byteCount the requested number of bytes
* @return the number of bytes actually read. Is less than bytesCount iff
* end of stream has been reached.
*/
virtual int read(char* begin, const int byteCount) override;
virtual int available() const override;
virtual bool isActive() const override;
virtual void close() override;
private:
Nan::Persistent<v8::Object> _stream;
bool _active;
JsEventLoopSync _evtLoop;
};
Of these functions, the important one here is read. The native library will call this function when it wants more data, and the function must block until it is able to return the requested data (or the stream ends). Here's my implementation of read:
int JsReadable2InputStream::read(char* begin, const int byteCount) {
if (!this->_active) { return 0; }
int read = -1;
while (read < 0 && this->_active) {
this->_evtLoop.invoke(
(voidLambda)[this,&read,begin,byteCount](){
v8::Local<v8::Object> stream = Nan::New(this->_stream);
const v8::Local<v8::Function> readFn = Nan::To<v8::Function>(Nan::Get(stream, JS_STR("read")).ToLocalChecked()).ToLocalChecked();
v8::Local<v8::Value> argv[] = { Nan::New<v8::Number>(byteCount) };
v8::Local<v8::Value> result = Nan::Call(readFn, stream, 1, argv).ToLocalChecked();
if (result->IsNull()) {
// Somewhat hacky/brittle way to check if stream has ended, but it's the only option
v8::Local<v8::Object> readableState = Nan::To<v8::Object>(Nan::Get(stream, JS_STR("_readableState")).ToLocalChecked()).ToLocalChecked();
if (Nan::To<bool>(Nan::Get(readableState, JS_STR("ended")).ToLocalChecked()).ToChecked()) {
// End of stream, all data has been read
this->_active = false;
read = 0;
return;
}
// Not enough data available, but stream is still open.
// Set a flag for the c++ thread to go to sleep
// This is the case that it gets stuck in
read = -1;
return;
}
v8::Local<v8::Object> bufferObj = Nan::To<v8::Object>(result).ToLocalChecked();
int len = Nan::To<int32_t>(Nan::Get(bufferObj, JS_STR("length")).ToLocalChecked()).ToChecked();
char* buffer = node::Buffer::Data(bufferObj);
if (len < byteCount) {
this->_active = false;
}
// copy the data out of the buffer
if (len > 0) {
std::memcpy(begin, buffer, len);
}
read = len;
}
);
if (read < 0) {
// Give js a chance to read more data
std::this_thread::sleep_for(std::chrono::milliseconds(10));
}
}
return read;
}
The idea is, the c++ code keeps a reference to the node stream object. When the native code wants to read, it has to synchronize with the node event loop, then attempt to invoke read on the node stream. If the node stream returns null, this indicates that the data isn't ready, so the native thread sleeps, giving the node event loop thread a chance to run and fill its buffers.
This solution works perfectly for a single stream, or even 2 or 3 streams running in parallel. Then for some reason when I hit the magical number of 4+ parallel streams, this totally deadlocks. None of the streams can successfully read any bytes at all. The above while loop runs infinitely, with the call into the node stream returning null every time.
It is behaving as though node is getting starved, and the streams never get a chance to populate with data. However, I've tried adjusting the sleep duration (to much larger values, and randomized values) and that had no effect. It is also clear that the event loop continues to run, since my lambda function continues to get executed there (I put some printfs inside to confirm this).
Just in case it might be relevant (I don't think it is), I'm also including my implementation of JsEventLoopSync. This uses libuv to schedule a lambda to be executed on the node event loop. It is designed such that only one can be scheduled at a time, and other invocations must wait until the first completes.
#include <nan.h>
#include <functional>
// simplified type declarations for the lambda functions
using voidLambda = std::function<void ()>;
// Synchronize with the node v8 event loop. Invokes a lambda function on the event loop, where access to js objects is safe.
// Blocks execution of the invoking thread until execution of the lambda completes.
class JsEventLoopSync {
public:
JsEventLoopSync() : _destroyed(false) {
// register on the default (same as node) event loop, so that we can execute callbacks in that context
// This takes a function pointer, which only works with a static function
this->_handles = new async_handles_t();
this->_handles->inst = this;
uv_async_init(uv_default_loop(), &this->_handles->async, JsEventLoopSync::_processUvCb);
// mechanism for passing this instance through to the native uv callback
this->_handles->async.data = this->_handles;
// mutex has to be initialized
uv_mutex_init(&this->_handles->mutex);
uv_cond_init(&this->_handles->cond);
}
~JsEventLoopSync() {
uv_mutex_lock(&this->_handles->mutex);
// prevent access to deleted instance by callback
this->_destroyed = true;
uv_mutex_unlock(&this->_handles->mutex);
// NOTE: Important, this->_handles must be a dynamically allocated pointer because uv_close() is
// async, and still has a reference to it. If it were statically allocated as a class member, this
// destructor would free the memory before uv_close was done with it (leading to asserts in libuv)
uv_close(reinterpret_cast<uv_handle_t*>(&this->_handles->async), JsEventLoopSync::_asyncClose);
}
// called from the native code to invoke the function
void invoke(const voidLambda& fn) {
if (v8::Isolate::GetCurrent() != NULL) {
// Already on the event loop, process now
return fn();
}
// Need to sync with the event loop
uv_mutex_lock(&this->_handles->mutex);
if (this->_destroyed) { return; }
this->_fn = fn;
// this will invoke processUvCb, on the node event loop
uv_async_send(&this->_handles->async);
// wait for it to complete processing
uv_cond_wait(&this->_handles->cond, &this->_handles->mutex);
uv_mutex_unlock(&this->_handles->mutex);
}
private:
// pulls data out of uv's void* to call the instance method
static void _processUvCb(uv_async_t* handle) {
if (handle->data == NULL) { return; }
auto handles = static_cast<async_handles_t*>(handle->data);
handles->inst->_process();
}
inline static void _asyncClose(uv_handle_t* handle) {
auto handles = static_cast<async_handles_t*>(handle->data);
handle->data = NULL;
uv_mutex_destroy(&handles->mutex);
uv_cond_destroy(&handles->cond);
delete handles;
}
// Creates the js arguments (populated by invoking the lambda), then invokes the js function
// Invokes resultLambda on the result
// Must be run on the node event loop!
void _process() {
if (v8::Isolate::GetCurrent() == NULL) {
// This is unexpected!
throw std::logic_error("Unable to sync with node event loop for callback!");
}
uv_mutex_lock(&this->_handles->mutex);
if (this->_destroyed) { return; }
Nan::HandleScope scope; // looks unused, but this is very important
// invoke the lambda
this->_fn();
// signal that we're done
uv_cond_signal(&this->_handles->cond);
uv_mutex_unlock(&this->_handles->mutex);
}
typedef struct async_handles {
uv_mutex_t mutex;
uv_cond_t cond;
uv_async_t async;
JsEventLoopSync* inst;
} async_handles_t;
async_handles_t* _handles;
voidLambda _fn;
bool _destroyed;
};
So, what am I missing? Is there a better way to wait for the node thread to get a chance to run? Is there a totally different design pattern that would work better? Does node have some upper limit on the number of streams that it can process at once?
As it turns out, the problems that I was seeing were actually client-side limitations. Browsers (and seemingly also node) have a limit on the number of open TCP connections to the same origin. I worked around this by spawning multiple node processes to do my testing.
If anyone is trying to do something similar, the code I shared is totally viable. If I ever have some free time, I might make it into a library.