manager::Memory not thread safe

[michaelsippel]

Now noticed there is already an issue for this #12

I ran into a memory error with the PMacc-GoL in a multithreaded case with resource manager.

After an investigation i found that in cuplaMallocHost(), the variable buf is a nullptr, which comes from manager::Memory::alloc(), which makes a reference out of an pointer (read from a map).
The variable manager::Memory::m_mapVector is not protected with a mutex, so there is a race condition indeed when doing multiple cuplaMallocHost() concurrently, causing one insert to be overwritten.
Could be solved with a simple mutex.

In cuplaMallocHost():

(gdb) select-frame 13
(gdb) info locals
extent = {static s_uiDim = <optimized out>, m_data = {8}}
buf = <error reading variable>
(gdb) p &buf
$2 = (alpaka::mem::buf::BufCpu<unsigned char, std::integral_constant<unsigned long, 1>, unsigned long> *) 0x0

So here is the relevant backtrace from my case:

#10 0x00005555556746da in
std::__shared_ptr_access<
    alpaka::mem::buf::cpu::detail::BufCpuImpl<
        unsigned char,
        std::integral_constant<unsigned long, 1ul>,
        unsigned long
    >,
    (__gnu_cxx::_Lock_policy)2,
    false,
    false
>::operator-> (this=0x0)
at /usr/include/c++/9.1.0/bits/shared_ptr_base.h:1015

#11 0x0000555555672c5e in
alpaka::mem::view::traits::GetPtrNative<
    alpaka::mem::buf::BufCpu<
        unsigned char,
        std::integral_constant<unsigned long, 1ul>,
    unsigned long
    >,
    void
>::getPtrNative
at .../alpaka/include/alpaka/mem/buf/BufCpu.hpp:291

#12 0x0000555555671279 in
alpaka::mem::view::getPtrNative<
    alpaka::mem::buf::BufCpu<
        unsigned char,
    std::integral_constant<unsigned long, 1ul>,
    unsigned long
    >
>
at .../alpaka/include/alpaka/mem/view/Traits.hpp:202

#13 0x000055555566f0ba in cupla_omp2_seq_async::cuplaMallocHost (ptrptr=0x5555558c0ba0, size=8)

[sbastrakov]

@michaelsippel thanks for reporting. Indeed, cupla API is currently not thread-safe and we need to some day fix it.

[sbastrakov]

@psychocoderHPC @ax3l do you think we should add a note to the readme (or somewhere) that cupla API is not thread safe, as was correctly pointed out in this issue?

[ax3l]

Yes, should be documented in the docs, including a few words what this implies/how to use it. (Just noting it might be too brief for non-computer scientists.)

[sbastrakov]

@ax3l okay, I could provide a PR tomorrow.