opal_memory_ptmalloc2_free() hanging when invalid address provided

[matcabral]

Testing PSM2 library, I found that deadlock is triggered when calling fee() with an invalid address. This is with OMPI 1.10.1. See stack trace below: It looks like opal_memory_ptmalloc2_free() locks a mutex before calling opal_memory_ptmalloc2_int_free(). Then, segfault occurs, the signal handler is called and we have opal_memory_ptmalloc2_free() called again inside the signal handler that will wait forever on the mutex locked by the first call.

#0  0x00002afb2bc7e99d in nanosleep () from /lib64/libpthread.so.0
#1  0x00002afb2c569005 in opal_memory_ptmalloc2_free ()
   from /usr/mpi/gcc/openmpi-1.10.0-hfi/lib64/libopen-pal.so.13
#2  0x00002afb2c4ee87f in opal_class_finalize ()
   from /usr/mpi/gcc/openmpi-1.10.0-hfi/lib64/libopen-pal.so.13
#3  0x00002afb2b782b5a in _dl_fini () from /lib64/ld-linux-x86-64.so.2
#4  0x00002afb2bec4e49 in __run_exit_handlers () from /lib64/libc.so.6
#5  0x00002afb2bec4e95 in exit () from /lib64/libc.so.6
#6  0x00002afb320fc2ca in hfi_sighdlr (sig=11, p1=<optimized out>,
ucv=<optimized out>)
    at opa_debug.c:190
#7  <signal handler called>
#8  0x00002afb2c568b2a in opal_memory_ptmalloc2_int_free ()
   from /usr/mpi/gcc/openmpi-1.10.0-hfi/lib64/libopen-pal.so.13
#9  0x00002afb2c569053 in opal_memory_ptmalloc2_free ()
   from /usr/mpi/gcc/openmpi-1.10.0-hfi/lib64/libopen-pal.so.13
#10 0x00002afb320edadc in ips_free_epaddr (epaddr=0x2afb34755540) at
ips_proto_connect.c:634
#11 ips_proto_disconnect (proto=proto@entry=0x1aeb180, force=force@entry=0,
numep=numep@entry=5160,
    array_of_epaddr=array_of_epaddr@entry=0x2afb3ea8a290,
    array_of_epaddr_mask=array_of_epaddr_mask@entry=0x2afb3ea80130,
    array_of_errors=array_of_errors@entry=0x2afb3ea851e0,
timeout_in=timeout_in@entry=52000000000)
    at ips_proto_connect.c:1439
#12 0x00002afb320e6e84 in ips_proto_fini (proto=proto@entry=0x1aeb180, force=0,
timeout_in=52000000000)
    at ips_proto.c:641
#13 0x00002afb320e001f in ips_ptl_fini (ptl=0x1aeb040, force=<optimized out>,
timeout_in=<optimized out>)
    at ptl.c:433
#14 0x00002afb320d36dd in __psm2_ep_close (ep=0x1aeac80, mode=0,
timeout_in=52000000000) at psm_ep.c:1107
#15 0x00002afb31ec0b17 in ompi_mtl_psm2_finalize ()
   from /usr/mpi/gcc/openmpi-1.10.0-hfi/lib64/openmpi/mca_mtl_psm2.so
#16 0x00002afb2b9dcd12 in ompi_mpi_finalize () from
/usr/mpi/gcc/openmpi-1.10.0-hfi/lib64/libmpi.so.12
#17 0x0000000000400d42 in main ()

[jsquyres]

@matcabral I think you should be using _exit() in your hfi_sighdlr() (instead of exit()). That might well resolve your issue...?

[matcabral]

Yes, thanks!
by definition _exit() should solve the hang. I haven't reproduced the hang so far. In any case, this is not an OMPI issue, so I'm closing this issue.