prov/shm: fix name compare bug

[aingerson]

Could result in a peer getting incorrectly unmmaped

[aingerson]

@shijin-aws I'm going to cherry pick this into 2.0.x. Want to put it on your radar because the 2.0.0 has this bug and it could have implications for you.

[shijin-aws]

@aingerson thanks! But I haven't seen any trouble in our side yet.... what is the impact of this bug?

[aingerson]

@shijin-aws We saw similar seg faults on close in Intel MPI that you reported on Open MPI
Similar to this one you reported:

Program terminated with signal SIGSEGV, Segmentation fault.
#0  0x00007f4547b07420 in smr_unmap_region (prov=0x7f4547bfb260 <smr_prov>, map=0x55c1c703b378, peer_id=0) at prov/shm/src/smr_util.c:506
506             if (peer_region->pid == getpid())
[Current thread is 1 (Thread 0x7f454d773fc0 (LWP 1545539))]
(gdb) bt
#0  0x00007f4547b07420 in smr_unmap_region (prov=0x7f4547bfb260 <smr_prov>, map=0x55c1c703b378, peer_id=0) at prov/shm/src/smr_util.c:506
#1  0x00007f4547b079e1 in smr_map_unmap (rbmap=0x55c1c703b398, node=0x55c1c715b490, context=0x0) at prov/shm/src/smr_util.c:601
#2  0x00007f4547b07acf in smr_map_del (map=0x55c1c703b378, shm_id=0) at prov/shm/src/smr_util.c:619
#3  0x00007f4547b04800 in smr_av_remove (av_fid=0x55c1c703b270, fi_addr=0x7f4537e33258, count=1, flags=0) at prov/shm/src/smr_av.c:220
#4  0x00007f4547a69a6b in fi_av_remove (av=0x55c1c703b270, fi_addr=0x7f4537e33258, count=1, flags=0) at ./include/rdma/fi_domain.h:531
#5  0x00007f4547a6d38b in efa_conn_rdm_deinit (av=0x55c1c70371e0, conn=0x7f4537e33220) at prov/efa/src/efa_av.c:358
#6  0x00007f4547a70a69 in efa_conn_release (av=0x55c1c70371e0, conn=0x7f4537e33220) at prov/efa/src/efa_av.c:555
#7  0x00007f4547a714b4 in efa_av_close_reverse_av (av=0x55c1c70371e0) at prov/efa/src/efa_av.c:794
#8  0x00007f4547a71591 in efa_av_close (fid=0x55c1c7037228) at prov/efa/src/efa_av.c:811
#9  0x00007f454c0c1000 in mca_btl_ofi_finalize () from /opt/amazon/openmpi/lib/openmpi/mca_btl_ofi.so
#10 0x00007f454d8fcddd in ?? () from /opt/amazon/openmpi/lib/libopen-pal.so.40
#11 0x00007f454d8e488c in mca_base_framework_close () from /opt/amazon/openmpi/lib/libopen-pal.so.40
#12 0x00007f454d8e488c in mca_base_framework_close () from /opt/amazon/openmpi/lib/libopen-pal.so.40
#13 0x00007f454dc46265 in ompi_mpi_finalize () from /opt/amazon/openmpi/lib/libmpi.so.40
#14 0x000055c1c603cd39 in main (argc=<optimized out>, argv=<optimized out>) at osu_get_acc_latency.c:115

[shijin-aws]

Oh yeah I call it out in the original PR, but somehow it doesn't show up any more so I let it go.

[aingerson]

@shijin-aws Yeah I thought we resolved it so I let it go too. Not sure why this only shows up in Intel MPI

[shijin-aws]

@aingerson you disabled intel mpi's shm in your test? We used to do that but recently only test with their default shm setting (using ofi:shm)

[aingerson]

@shijin-aws I'm not sure, the bug report came from IMPI but they must have disabled their internal shm to get here. If you only use IMPI shm for IMPI then this shouldn't be a big issue for you! But if you see any segfaults on finalize anywhere else then it's probably this