Omit the first check of the double-checked locking pattern in recordException() in parallel.cpp when CV_THREAD_SANITIZER is defined. This should only slow recordException() down when the thread sanitizer is used, and avoids the TSAN data race warning.

CV_THREAD_SANITIZER is defined in cvdef.h which is included by core.hpp which is included by utility.hpp which is included by precomp.hpp which is included by parallel.cpp. If this is too indirect, #include "cvdef.h" can be added in parallel.cpp.

Original report

Clang's thread sanitizer found the following when running an internal fuzz target:

WARNING: ThreadSanitizer: data race (pid=8714)
Read of size 1 by thread T6:
    #0 cv::(anonymous namespace)::ParallelLoopBodyWrapperContext::recordException() OpenCV/modules/core/src/parallel.cpp:278
    #1 cv::(anonymous namespace)::ParallelLoopBodyWrapper::operator()(cv::Range const&) const third_party/OpenCV/modules/core/src/parallel.cpp:357
Previous write of size 1 at 0x7ffc7c20ee58 by main thread (mutexes: write M0):
    #0 cv::(anonymous namespace)::ParallelLoopBodyWrapperContext::recordException() third_party/OpenCV/modules/core/src/parallel.cpp:283
    #1 cv::(anonymous namespace)::ParallelLoopBodyWrapper::operator()(cv::Range const&) const third_party/OpenCV/modules/core/src/parallel.cpp:357

Location is stack of main thread.
Location is global '??'

Mutex M0 created at:
    #3 cv::getInitializationMutex() third_party/OpenCV/modules/core/src/system.cpp:93
Thread T6 (running) created by main thread at:
    #1 cv::WorkerThread::WorkerThread(cv::ThreadPool&, unsigned int) third_party/OpenCV/modules/core/src/parallel_impl.cpp:241

SUMMARY: ThreadSanitizer: data race third_party/OpenCV/modules/core/src/parallel.cpp:278 in cv::(anonymous namespace)::ParallelLoopBodyWrapperContext::recordException()

I believe it is a false positive in this case. As long as pException is not accessed until all threads are joined (in finalize() only), I do not see how the lack of order guarantee in the body of the second !hasException condition is an issue (unlike the anti-pattern DCLP singleton initialization).

In other words, I think the code at head works as intended as long as C++ guarantees the following, and I guess it does:

• The second if (!hasException) is not optimized out by the first if (!hasException)
• std::lock_guard happens before the second if (!hasException)
• The second if (!hasException) happens before hasException = true;
• The second if (!hasException) happens before exception = std::current_exception();

TSAN may be triggered because it detects a read during a write, or maybe it sees a branching on the first if (!hasException) that differs depending on the run. But the data race here has no effect: it is the purpose of the double-checked locking pattern.

Alternatives

The current solution is simple but I considered other options. Feel free to pick any or suggest another.

• Just remove the first check all the time because recordException() is not performance-critical.
• std::call_once():

  bool hasException;
  std::once_flag flagException;
  std::exception_ptr pException;
  void recordException()
  {
      std::call_once(flagException, [this] {
          hasException = true;
          pException = std::current_exception();
      });
  }

• Something similar to https://preshing.com/20130930/double-checked-locking-is-fixed-in-cpp11/

Related

• Replaced incorrect double checked locking with a static #21612
• Fix two race conditions in LazyInit and Matrix default allocator #8713
• core(parallel): suppress TSAN warning #20803
• core(tls): avoid destruction of TlsAbstraction singleton #20841
• Race condition in TlsAbstraction #20606

Pull Request Readiness Checklist

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

• I agree to contribute to the project under Apache 2 License.
• To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
• The PR is proposed to the proper branch
• There is a reference to the original bug report and related work