dash::copy broken on team-allocations

[devreal]

Consider the following example code:

TEST_F(CopyTest, BlockingGlobalToLocalTeam)
{
  // Copy all elements of global array into local vector:
  const int num_elem_per_unit = 20;
  size_t num_elem_total       = _dash_size * num_elem_per_unit;

  auto& team = dash::Team::All().split(2);

  dash::Array<int> array(num_elem_total, dash::BLOCKED, team);

  // Assign initial values: [ 1000, 1001, 1002, ... 2000, 2001, ... ]
  for (auto l = 0; l < num_elem_per_unit; ++l) {
    array.local[l] = ((dash::myid() + 1) * 1000) + l;
  }
  array.barrier();

  // Local vector to store copy of global array;
  int *local_vector = new int[num_elem_per_unit * team.size()];

  // Copy values from global range to local memory.
  // All units copy first block, so unit 0 tests local-to-local copying.
  dash::copy(array.begin(), array.end(), local_vector);

  for (size_t i = 0; i < array.size(); ++i) {
    EXPECT_EQ_U(static_cast<int>(array[i]), local_vector[i]);
  }

  delete[] local_vector;
}

Executing the test on 4 units, Valgrind reports the following invalid write:

==1559== Invalid write of size 2
==1559==    at 0x4C32723: memcpy@@GLIBC_2.14 (in /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
==1559==    by 0x6B06205: opal_convertor_unpack (in /home/joseph/opt/openmpi-2.1.0/lib/libopen-pal.so.20.10.0)
==1559==    by 0xCD399C5: mca_pml_ob1_recv_frag_callback_match (in /home/joseph/opt/openmpi-2.1.0/lib/openmpi/mca_pml_ob1.so)
==1559==    by 0xB8CDD09: mca_btl_vader_component_progress (in /home/joseph/opt/openmpi-2.1.0/lib/openmpi/mca_btl_vader.so)
==1559==    by 0x6AF574B: opal_progress (in /home/joseph/opt/openmpi-2.1.0/lib/libopen-pal.so.20.10.0)
==1559==    by 0x6AFA824: sync_wait_mt (in /home/joseph/opt/openmpi-2.1.0/lib/libopen-pal.so.20.10.0)
==1559==    by 0x52A5D4B: ompi_request_default_wait (in /home/joseph/opt/openmpi-2.1.0/lib/libmpi.so.20.10.0)
==1559==    by 0x52DF46C: PMPI_Wait (in /home/joseph/opt/openmpi-2.1.0/lib/libmpi.so.20.10.0)
==1559==    by 0xA35F20: dart_get_blocking (dart_communication.c:970)
==1559==    by 0x7D47DC: int* dash::internal::copy_impl<int, dash::GlobIter<int, dash::BlockPattern<1, (dash::MemArrange)1, long>, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> >, dash::GlobPtr<int, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> > >, dash::GlobRef<int> > >(dash::GlobIter<int, dash::BlockPattern<1, (dash::MemArrange)1, long>, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> >, dash::GlobPtr<int, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> > >, dash::GlobRef<int> >, dash::GlobIter<int, dash::BlockPattern<1, (dash::MemArrange)1, long>, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> >, dash::GlobPtr<int, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> > >, dash::GlobRef<int> >, int*) (Copy.h:178)
==1559==    by 0x7CE128: int* dash::copy<int, dash::GlobIter<int, dash::BlockPattern<1, (dash::MemArrange)1, long>, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> >, dash::GlobPtr<int, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> > >, dash::GlobRef<int> > >(dash::GlobIter<int, dash::BlockPattern<1, (dash::MemArrange)1, long>, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> >, dash::GlobPtr<int, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> > >, dash::GlobRef<int> >, dash::GlobIter<int, dash::BlockPattern<1, (dash::MemArrange)1, long>, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> >, dash::GlobPtr<int, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> > >, dash::GlobRef<int> >, int*) (Copy.h:1002)
==1559==    by 0x7C860D: CopyTest_BlockingGlobalToLocalTeam_Test::TestBody() (CopyTest.cc:858)
==1559==  Address 0x1890fb48 is 0 bytes after a block of size 56 alloc'd
==1559==    at 0x4C2E0EF: operator new(unsigned long) (in /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
==1559==    by 0xA1A79B: dash::util::LocalityDomain::init(dart_domain_locality_s*) (LocalityDomain.cc:603)
==1559==    by 0xA16946: dash::util::LocalityDomain::LocalityDomain(dart_domain_locality_s*) (LocalityDomain.cc:67)
==1559==    by 0x790260: dash::util::UnitLocality::UnitLocality(dash::Team const&, dash::unit_id<(dash::unit_scope)0, dart_local_unit>) (UnitLocality.h:62)
==1559==    by 0x7CD8BD: int* dash::copy<int, dash::GlobIter<int, dash::BlockPattern<1, (dash::MemArrange)1, long>, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> >, dash::GlobPtr<int, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> > >, dash::GlobRef<int> > >(dash::GlobIter<int, dash::BlockPattern<1, (dash::MemArrange)1, long>, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> >, dash::GlobPtr<int, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> > >, dash::GlobRef<int> >, dash::GlobIter<int, dash::BlockPattern<1, (dash::MemArrange)1, long>, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> >, dash::GlobPtr<int, dash::GlobStaticMem<int, dash::allocator::SymmetricAllocator<int> > >, dash::GlobRef<int> >, int*) (Copy.h:847)
==1559==    by 0x7C860D: CopyTest_BlockingGlobalToLocalTeam_Test::TestBody() (CopyTest.cc:858)
==1559==    by 0xA0967E: void testing::internal::HandleSehExceptionsInMethodIfSupported<testing::Test, void>(testing::Test*, void (testing::Test::*)(), char const*) (in /home/joseph/src/dash/dash/build/dash/dash-test-mpi)
==1559==    by 0xA0371E: void testing::internal::HandleExceptionsInMethodIfSupported<testing::Test, void>(testing::Test*, void (testing::Test::*)(), char const*) (in /home/joseph/src/dash/dash/build/dash/dash-test-mpi)
==1559==    by 0x9E7DCD: testing::Test::Run() (in /home/joseph/src/dash/dash/build/dash/dash-test-mpi)
==1559==    by 0x9E8765: testing::TestInfo::Run() (in /home/joseph/src/dash/dash/build/dash/dash-test-mpi)
==1559==    by 0x9E8E58: testing::TestCase::Run() (in /home/joseph/src/dash/dash/build/dash/dash-test-mpi)
==1559==    by 0x9EFF9F: testing::internal::UnitTestImpl::RunAllTests() (in /home/joseph/src/dash/dash/build/dash/dash-test-mpi)

Eventually the process goes down in a Segfault.

Interestingly, leaving out the team split (.split(2)) leads to correct results so I assume that dash::copy does not honor the team in the array.

Related: The failing ThreadsafetyTest.ConcurrentAlgorithm mentioned by @ddiefenthaler in #292

[devreal]

Never mind, the test case was faulty (allocated _dash_size * num_elem_per_unit instead of team.size() * num_elem_per_unit element.