Snap for 6793014 from 07602cda to master

    mGpuMem->traverseGpuMemTotals([](int64_t ts, uint32_t gpuId, uint32_t pid, uint64_t size) {

        GpuMemDataSource::Trace([&](GpuMemDataSource::TraceContext ctx) {

            auto packet = ctx.NewTracePacket();

            packet->set_timestamp_clock_id(perfetto::protos::pbzero::BUILTIN_CLOCK_MONOTONIC);

            packet->set_timestamp(ts);

            auto* event = packet->set_gpu_mem_total_event();

            event->set_gpu_id(gpuId);

void CallbackScheduler::loop() {

    while (true) {

        std::unique_lock<std::mutex> lock(mMutex);

        std::lock_guard<std::mutex> lock(mMutex);

        if (mFinished) {

            // Destructor was called, so let the callback thread die.

            break;

        }

        while (!mQueue.empty() && mQueue.top().isExpired()) {

            DelayedCallback callback = mQueue.top();

            mQueue.top().run();

            mQueue.pop();

            lock.unlock();

            callback.run();

            lock.lock();

        }

        if (mQueue.empty()) {

            // Wait until a new callback is scheduled.