Re: Bug#848574: openmpi: frequent segfault in linker on mips64el

James Cowgill <[email protected]> Sun, 18 Dec 2016 17:21:00 +0000

Newsgroups	gmane.linux.debian.ports.mips
Message-ID	<[email protected]>

Hi,

On 18/12/16 13:44, Ansgar Burchardt wrote:
> Package: libopenmpi-dev
> Version: 2.0.2~git.20161225-8
> Severity: important
> 
> [ CC'ed debian-mips@ as a segfault in the linker might not be a bug in MPI ]
> 
> OpenMPI crashed frequenly in the runtime linker on mips64el, even for a
> trivial program calling just MPI_Init and MPI_Finalize:
> 
> +---
> | #include <mpi.h>
> |
> | int main(int argc, char** argv)
> | {
> |   MPI_Init(&argc, &argv);
> |   MPI_Finalize();
> |   return 0;
> | }
> +---[ test.cc ]
> 
> After building with `mpic++ -o test test.cc`, I often get the following
> backtrace from the core file:
> 
> +---
> | Core was generated by `./test'.
> | Program terminated with signal SIGSEGV, Segmentation fault.
> | #0  elf_machine_runtime_link_map (gpreg=<optimized out>, stub_pc=1099310712688)
> |     at ../sysdeps/mips/dl-trampoline.c:84
> | 84	../sysdeps/mips/dl-trampoline.c: No such file or directory.
> | [Current thread is 1 (LWP 2837)]
> | (gdb) bt
> | #0  elf_machine_runtime_link_map (gpreg=<optimized out>, stub_pc=1099310712688)
> |     at ../sysdeps/mips/dl-trampoline.c:84
> | #1  __dl_runtime_resolve (sym_index=151, return_address=<optimized out>, old_gpreg=<optimized out>, 
> |     stub_pc=1099310712688) at ../sysdeps/mips/dl-trampoline.c:125
> | #2  0x000000fff5e341ac in _dl_runtime_resolve () from /lib64/ld.so.1
> | Backtrace stopped: frame did not save the PC
> +---[ backtrace from `gdb ./test core` ]
> 
> I noticed as src:dune-common/2.5.0-1 failed on mips64el even though the last
> upload build sucessfully.

This symptoms here look very much like those in the DT_MIPS_RLD_MAP_REL
bug from #818909. If "mpic++" simply calls g++ with some extra args (ie
doesn't do anything to the executable after ld has touched it), then I
think binutils is the likely culprit here.

I'm not sure why it only happens on mips64el though (instead of all 3
mips arches).

[sigh, I swear we didn't use to have this many bugs in the mips toolchain]

Thanks,
James

signature.asc (application/pgp-signature, 833 B)