Unit3 deadlock

[Chris Beall]

Unit3 sometimes deadlocks. On Ubuntu 14.04, in Debug mode + TBB this happens roughly 1/10 times.

I noticed because Jenkins was unable to get through its latest build. I just configured a unit test timeout on Jenkins so it will count this as a failure and move on.

[Hayk Martirosyan]

Oh man, that's a bummer. I moved the mutex from a global to a member variable. I recall running into this issue at some point, and then before I checked in I tried to reproduce it many times and couldn't get anything, so I thought it had been a red herring. Didn't realize it only happened in debug mode. Frank and I looked at it together back then and determined that it was properly used. Though I am used to the C++11 variants, so maybe there's a difference there with the TBB? I recall trying to make it recursive to see if that would fix the issue, but as far as I can tell nothing in basis should be invoking basis, because that would be bad for several reasons.

[Chris Beall]

Here's a hint. Compiling this on Mac gives this error:

/Users/cbeall3/git/gtsam/gtsam/geometry/Unit3.h:156:12: error: call to implicitly-deleted copy constructor of 'gtsam::Unit3'
    return Unit3(p_.cross(q.p_));
           ^~~~~~~~~~~~~~~~~~~~~
/Users/cbeall3/git/gtsam/gtsam/geometry/Unit3.h:51:22: note: copy constructor of 'Unit3' is implicitly deleted because field
      'B_mutex_' has a deleted copy constructor
  mutable tbb::mutex B_mutex_; ///< Mutex to protect the cached basis.
                     ^
/opt/local/include/tbb/mutex.h:40:15: note: copy constructor of 'mutex' is implicitly deleted because base class
      'internal::mutex_copy_deprecated_and_disabled' has a deleted copy constructor
class mutex : internal::mutex_copy_deprecated_and_disabled {
              ^
/opt/local/include/tbb/tbb_stddef.h:334:44: note: copy constructor of 'mutex_copy_deprecated_and_disabled' is implicitly deleted
      because base class 'tbb::internal::no_copy' has an inaccessible copy constructor
class mutex_copy_deprecated_and_disabled : no_copy {};
                                           ^

[Frank Dellaert]

I just built current develop in Debug mode on my mac - no such error :-(

[Frank Dellaert]

A possible fix is to add an explicit copy constructor, which default-constructs the mutex, copies only p_.

[Frank Dellaert]

Since you, @cbeall3 , are the only one who has that error (puzzling!) seems like a quick thing to try?

[Chris Beall]

Must be compiler/tbb version. I'm on the very latest:
XCode 7.1 with Clang Apple LLVM version 7.0.0 (clang-700.1.76) and tbb 4.4-20150728_0.

[Frank Dellaert]

My clang is Apple LLVM version 7.0.0 (clang-700.1.76)

How do I get TBB version?

[Chris Beall]

sudo port installed | grep tbb

[Frank Dellaert]

Fixed in PR #194