master: "make check" unpack_hetero datatype test segv

[jsquyres]

When configured statically on a Linux x86_64 machine, make check fails the unpack_hetero test with a segv. Here's the backtrace:

(gdb) bt
#0  0x000000000040fd7c in opal_datatype_compute_ptypes (datatype=0x5810c0 <opal_datatype_int4>) at opal_datatype_get_count.c:163
#1  0x000000000040b714 in opal_datatype_compute_remote_size (pData=0x5810c0 <opal_datatype_int4>, sizes=0x7f48c8) at opal_convertor.c:456
#2  0x000000000040b82f in opal_convertor_compute_remote_size (pConvertor=0x7f4760) at opal_convertor.c:483
#3  0x000000000040ba3e in opal_convertor_prepare_for_recv (convertor=0x7f4760, datatype=0x5810c0 <opal_datatype_int4>, count=1, pUserBuf=0x7fffffffca40) at opal_convertor.c:569
#4  0x0000000000405d3d in main (argc=1, argv=0x7fffffffcb68) at unpack_hetero.c:61

The code in question is:

163 =>  datatype->ptypes = (size_t*)calloc(OPAL_DATATYPE_MAX_SUPPORTED, sizeof(size_t));

gdb shows:

(gdb) p datatype
$2 = (opal_datatype_t *) 0x5810c0 <opal_datatype_int4>
(gdb) p datatype->ptypes
$3 = (size_t *) 0x0

I'm not quite sure why this is a segv. Is this read-only member, perchance?

[jsquyres]

Forgot to mention -- here's the MTT showing where it is happening (and I can reproduce manually): https://mtt.open-mpi.org/index.php?do_redir=2442

[ggouaillardet]

@jsquyres yes, this is likely the root cause

OPAL_DECLSPEC const opal_datatype_t opal_datatype_int4 = OPAL_DATATYPE_INITIALIZER_INT4(0);

[jsquyres]

@ggouaillardet What should it be?

[ggouaillardet]

@jsquyres

with shared libs

61	    if( OPAL_SUCCESS != opal_convertor_prepare_for_recv( pConv, &opal_datatype_int4, 1, rbuf ) ) {
(gdb) p &opal_datatype_int4
$1 = (const opal_datatype_t *) 0x6020a0 <opal_datatype_int4>

0000000000602000      4K rw--- unpack_hetero

but with static libs

61	    if( OPAL_SUCCESS != opal_convertor_prepare_for_recv( pConv, &opal_datatype_int4, 1, rbuf ) ) {
(gdb) p &opal_datatype_int4
$6 = (const opal_datatype_t *) 0x57fc80 <opal_datatype_int4>

0000000000400000   1952K r-x-- unpack_hetero

removing the const qualifier is an option, but i suspect it is an overkill
i'll try to see if i can improve that (by not computing ptypes for a predefined datatype (since that should be pretty trivial)

[jsquyres]

@bosilca says he's looking at this.