Fix #505: remove data race in threaded CSV parser

[hcho3]

Minimal example to detect data race:

// example.cc
#include <dmlc/registry.h>
#include <dmlc/data.h>
#include <cstdint>

int main(void) {
  std::unique_ptr<dmlc::Parser<uint32_t, float>> parser(
    dmlc::Parser<uint32_t, float>::Create("data.txt", 0, 1, "csv"));
  parser->BeforeFirst();
  size_t total = 0;
  while (parser->Next()) {
    const auto& value = parser->Value();
    LOG(INFO) << "value.size = " << value.size;
    total += value.size;
  }
  LOG(INFO) << "total = " << total;

  return 0;
}

Content of data.txt: data.txt

To run the example:

1. Checkout the latest master branch of dmlc-core.
2. Set CFLAGS in Makefile of dmlc-core as follows:

# inside Makefile
export CFLAGS = -g -O0 -fsanitize=thread -Wall -Wno-unknown-pragmas -Iinclude

3. Compile dmlc-core by invoking make at the project root of dmlc-core
4. Compile example.cc by linking against dmlc-core:

g++ -g -O0 -fsanitize=thread -o example example.cc ./dmlc-core/libdmlc.a \
  -I./dmlc-core/include -lpthread -lgomp

You should see a long error trace, consisting of many occurrences of data race.

The proposed changes will remove all data race errors.

[szha]

Thanks for the fix! How does the speed look like before and after the change? This fix is necessary regardless, so the question is mostly out of curiosity.

[hcho3]

@szha I'll try to grab a large CSV file and do a performance test.