free() invalid pointer NewtonSolver destructor

[lmolin3]

free() invalid pointer NewtonSolver destructor

[v-dobrev]

This looks like memory corruption from somewhere else, not from NewtonSolver. Can you try building MFEM in debug mode (option MFEM_DEBUG=YES) and run your code? In debug mode, the library performs some additional checks which may catch the issue. If this build runs with the same error, try running the code under valgrind -- that can also help you find memory related issues.

[lmolin3]

You were right, the valgrind one was an issue with the MPI implementation. Apparently with the docker image I was working on had Intel MPI (causing troubles), while OpenMPI v4.1 works.
Thanks for the suppression file as well, it'll help spotting the error in the code.

[jandrej]

https://godbolt.org/z/8a3T6rEfa

works as expected

[lmolin3]

I know the NewtonSolver is not the problem, I've used it before and didn't have any problems. I was wondering If you have any idea of where the problem may come from in this case. Thanks again!

[lmolin3]

@v-dobrev Appreciate your help! I was already working in debug mode. I've experienced issues in the past using valgrind on MFEM in debug mode, but installing valgrind.3.22.0 seemed to solve that issue.

Running it on one of the parallel examples (ex1p) gives signal SIGILL on MPI::Init(), am I doing anything wrong?

mpirun -np 2 valgrind --tool=memcheck --verbose --trace-children=yes --leak-check=full --show-reachable=yes --track-
origins=yes --log-file=/home/euler/develop/valgrind-logs/test.log ./ex1p

==99063==    at 0x65D7932: hwloc_list_special_objects (in /usr/local/lib/libhwloc.so.15.6.2)
==99063==    by 0x65D7F75: hwloc_connect_special_levels (in /usr/local/lib/libhwloc.so.15.6.2)
==99063==    by 0x65DDBA2: hwloc_topology_reconnect (in /usr/local/lib/libhwloc.so.15.6.2)
==99063==    by 0x65E2AA4: hwloc_topology_load (in /usr/local/lib/libhwloc.so.15.6.2)
==99063==    by 0x5123A1A: MPII_hwtopo_init (in /usr/local/lib/libmpi.so.12.3.0)
==99063==    by 0x50B5AB7: MPII_Init_thread (in /usr/local/lib/libmpi.so.12.3.0)
==99063==    by 0x50B577E: MPIR_Init_impl (in /usr/local/lib/libmpi.so.12.3.0)
==99063==    by 0x4EC23A8: internal_Init (in /usr/local/lib/libmpi.so.12.3.0)
==99063==    by 0x4EC2445: PMPI_Init (in /usr/local/lib/libmpi.so.12.3.0)
==99063==    by 0x1690EE: mfem::Mpi::Init_(int*, char***) (source/mfem/mesh/../general/communication.hpp:79)
==99063==    by 0x1675FF: mfem::Mpi::Init() (source/mfem/mesh/../general/communication.hpp:36)
==99063==    by 0x165E4D: main (source/mfem/examples/ex1p.cpp:72)

[v-dobrev]

My suspicion is that the valgrind + MPI issue is not mfem related. Can you try with a simple "hello world"-like MPI program?

We do nightly tests with valgrind + MPI and they do not show any issues. We do need to use a suppression file for some MPI-related issues and older hypre versions. Here's what it looks like:

{
   MPI-related-write
   Memcheck:Param
   writev(vector[...])
   fun:writev
}

{
   MPI-related-syscall
   Memcheck:Param
   sched_setaffinity(mask)
   fun:syscall
}

{
   MPI_Init-memcpy
   Memcheck:Overlap
   fun:memcpy
   fun:pmgr_allgather
   fun:MPID_PSM_Init
   fun:MPID_Init
   fun:MPIR_Init
   fun:main
}

{
   OMPI_Init-param-getsockopt-optlen
   Memcheck:Param
   socketcall.getsockopt(optlen)
   ...
   fun:ompi_mpi_init
   fun:PMPI_Init*
}

{
   OMPI_Init-param-getsockopt-optlen_out
   Memcheck:Param
   socketcall.getsockopt(optlen_out)
   ...
   fun:ompi_mpi_init
   fun:PMPI_Init*
}

{
   OMPI_Init-param-setsockopt-optlen
   Memcheck:Param
   setsockopt(optlen)
   ...
   fun:ompi_mpi_init
   fun:PMPI_Init*
}

{
   OMPI_Init-leak
   Memcheck:Leak
   ...
   fun:ompi_mpi_init
}

{
   OMPI_Finalize-leak
   Memcheck:Leak
   ...
   fun:ompi_mpi_finalize
}

{
   hypre-BoomerAMG-GaussElimSetup
   Memcheck:Overlap
   fun:memcpy
   fun:ompi_ddt_copy_content_same_ddt
   fun:ompi_coll_tuned_allgather_intra_bruck
   fun:PMPI_Allgather
   fun:hypre_GaussElimSetup
   fun:hypre_BoomerAMGSetup
}

{
   hypre-BoomerAMG_GMExpandInterp
   Memcheck:Leak
   match-leak-kinds: definite
   fun:calloc
   fun:hypre_CAlloc
   fun:hypre_BoomerAMG_GMExpandInterp
   fun:hypre_BoomerAMGSetup
}

What MPI implementation are you using? I had some issues with OpenMPI v5.0.x but v4.1.x should work fine. I'm not sure about MPICH, or other implementations.