if (dst_fd < 0) {
    close(src_fd);
    return -1;
  }

  // Both files have been opened, do the transfer.
  // Since errno would be overridden by `close` below, save it here.
  int error_code = 0;
  if (CopyFileContents(dst_fd, src_fd, size) < 0) error_code = errno;

  close(src_fd);
  close(dst_fd);
  if (error_code != 0) {
    errno = error_code;
    return -1;
  } else {

// In the TFv1 convention, TF_Tensor can hold a string serialization of
// DT_RESOURCE. The string serialization is converted back to a
// ResourceHandle during Session run where the TF_Tensor is converted to a
// Tensor.
// TFv2 does not depend on this conversion. There is no matching
// TF_TensorFromTensorV1 because the conversion to string is performed by the
// python side of Session.

    return;
  } else if (S_ISDIR(st.st_mode)) {
    TF_SetStatus(status, TF_FAILED_PRECONDITION, "source path is a directory");
    return;
  }

  // Do the actual rename. Here both arguments are filenames.
  if (rename(src, dst) != 0)
    TF_SetStatusFromIOError(status, errno, dst);
  else
    TF_SetStatus(status, TF_OK, "");
}

  // threadpool, so that we avoid the possibility of running the runner_ in the
  // threadpool of GPU event mgr, as that can trigger more callbacks to be
  // scheduled on that same threadpool, causing a deadlock in cases where the
  // caller of event_mgr->ThenExecute() blocks on the completion of the callback
  // (as in the case of ConstOp kernel creation on GPU, which involves copying a
  // CPU tensor to GPU).

  }

  dlm_tensor->dl_tensor.shape = shape_arr->data();
  // There are two ways to represent compact row-major data
  // 1) nullptr indicates tensor is compact and row-majored.
  // 2) fill in the strides array as the real case for compact row-major data.
  // Here we choose option 2, since some frameworks didn't handle the strides
  // argument properly.

    ASSERT_TF_OK(s) << "TF_GetChildren failed for " << dirpath;
    const char* childpath;
    ASSERT_TRUE(TF_StringStreamNext(children, &childpath));
    ASSERT_EQ(::tensorflow::string(childpath), "somefile.txt");
    // There should only be one file in this directory.
    ASSERT_FALSE(TF_StringStreamNext(children, &childpath));
    ASSERT_EQ(childpath, nullptr);
    TF_StringStreamDone(children);

    TF_FileStatistics stats;

};

TEST_P(CppGradients, TestIdentityNGrad) {
  // This test is interesting because the current implementation of GradientTape
  // would return [0, 1] whereas we use build_default_zeros_grads=false here
  // so we get back [nullptr, 1].

  AbstractTensorHandlePtr x1;
  {
    AbstractTensorHandle* x1_raw = nullptr;
    status_ = TestScalarTensorHandle<float, TF_FLOAT>(

  void MarkAsResult(AbstractTensorHandle* gradient) const override;

  void DeleteGradient(AbstractTensorHandle* gradient) const override;

 private:
  // The context where the aggregation op `Add` is to be created.
  AbstractContext* ctx_;
};

// Returns the number of elements in the gradient tensor.
int64_t TapeVSpace::NumElements(AbstractTensorHandle* tensor) const {

  TF_AddInput(desc, {zero, 0});
  TF_AddInput(desc, {input, 0});
  TF_SetAttrInt(desc, "num_split", 3);
  TF_SetAttrType(desc, "T", TF_INT32);
  // Set device to CPU since there is no version of split for int32 on GPU
  // TODO(iga): Convert all these helpers and tests to use floats because
  // they are usually available on GPUs. After doing this, remove TF_SetDevice
  // call in c_api_function_test.cc

void RamFileBlockCache::Trim() {
  while (!lru_list_.empty() && cache_size_ > max_bytes_) {
    RemoveBlock(block_map_.find(lru_list_.back()));
  }
}

/// Move the block to the front of the LRU list if it isn't already there.
void RamFileBlockCache::UpdateLRU(const Key& key,
                                  const std::shared_ptr<Block>& block,
                                  TF_Status* status) {
  absl::MutexLock lock(&mu_);