System librocalution can cause issue due to implicit loading of UCX

[vchuravy]

Originally noted by @leios in JuliaMolSim/Molly.jl#147 (comment)

I was trying to run some code on my AMD laptop, and I was experiencing bizarre errors that were stemming from UCS/UCX.
Very similar to https://juliaparallel.org/MPI.jl/stable/knownissues/#UCX, but surprising since I was not using MPI.

using AMDGPU

import Libdl
for dll in Libdl.dllist()
    @show dll
end

Showed:

dll = "/opt/rocm/lib/libhsakmt.so.1"
dll = "/usr/lib/libdrm.so.2"
dll = "/usr/lib/libdrm_amdgpu.so.1"
dll = "/usr/lib/libnuma.so.1"
dll = "/opt/rocm/lib/libamdhip64.so"
dll = "/opt/rocm/lib/libamd_comgr.so.2"
dll = "/opt/rocm/lib/libhsa-runtime64.so.1"
dll = "/usr/lib/libzstd.so.1"
dll = "/usr/lib/libncursesw.so.6"
dll = "/usr/lib/libelf.so.1"
dll = "/usr/lib/libucs.so.0"
dll = "/usr/lib/libucm.so.0"
dll = "/usr/lib/libsframe.so.1"

Using https://github.com/haampie/libtree

libtree /opt/rocm/lib/*
librocalution.so.1
├── libomp.so [runpath]
├── libmpi.so.40 [runpath]
│   ├── libopen-pal.so.80 [runpath]
│   │   ├── libfabric.so.1 [runpath]
│   │   │   └── libnuma.so.1 [default path]
│   │   ├── libpmix.so.2 [runpath]
│   │   │   ├── libhwloc.so.15 [ld.so.conf]
│   │   │   ├── libevent_core-2.1.so.7 [default path]
│   │   │   └── libevent_pthreads-2.1.so.7 [default path]
│   │   │       └── libevent_core-2.1.so.7 [default path]
│   │   ├── libhwloc.so.15 [runpath]
│   │   │   └── libudev.so.1 [default path]
│   │   │       └── libcap.so.2 [default path]
│   │   ├── libevent_pthreads-2.1.so.7 [runpath]
│   │   ├── libevent_core-2.1.so.7 [runpath]
│   │   ├── libuct.so.0 [runpath]
│   │   │   └── libucs.so.0 [default path]
│   │   │       ├── libzstd.so.1 [default path]
│   │   │       ├── libz.so.1 [default path]
│   │   │       ├── libsframe.so.1 [default path]
│   │   │       └── libucm.so.0 [default path]
│   │   ├── libucm.so.0 [runpath]
│   │   ├── libucs.so.0 [runpath]
│   │   └── libucp.so.0 [runpath]
│   │       ├── libuct.so.0 [default path]
│   │       └── libucs.so.0 [default path]
│   ├── libhwloc.so.15 [runpath]
│   ├── libevent_pthreads-2.1.so.7 [runpath]
│   ├── libevent_core-2.1.so.7 [runpath]
│   ├── libpmix.so.2 [runpath]
│   ├── libucs.so.0 [runpath]
│   ├── libucp.so.0 [runpath]
│   └── libfabric.so.1 [runpath]
└── librocalution_hip.so.1 [ld.so.conf]
    ├── librocblas.so.4 [ld.so.conf]
    │   └── libamdhip64.so.6 [ld.so.conf]
    │       ├── libamd_comgr.so.2 [ld.so.conf]
    │       │   ├── libz.so.1 [default path]
    │       │   ├── libncursesw.so.6 [default path]
    │       │   └── libzstd.so.1 [default path]
    │       ├── libhsa-runtime64.so.1 [ld.so.conf]
    │       │   ├── libhsakmt.so.1 [ld.so.conf]
    │       │   │   ├── libdrm.so.2 [default path]
    │       │   │   ├── libnuma.so.1 [default path]
    │       │   │   └── libdrm_amdgpu.so.1 [default path]
    │       │   │       └── libdrm.so.2 [default path]
    │       │   ├── libdrm.so.2 [default path]
    │       │   └── libelf.so.1 [default path]
    │       │       ├── libz.so.1 [default path]
    │       │       └── libzstd.so.1 [default path]
    │       └── libnuma.so.1 [default path]
    ├── libamdhip64.so.6 [ld.so.conf]
    ├── librocrand.so.1 [ld.so.conf]
    │   └── libamdhip64.so.6 [ld.so.conf]
    └── librocsparse.so.1 [ld.so.conf]
        └── libamdhip64.so.6 [ld.so.conf]

So on my system librocalution is pulling in OpenMPI and thus loads UCX implicitly.

[vchuravy]

On my system this looked like:

 dev [loki:132223:0:132223] Caught signal 11 (Segmentation fault: invalid permissions for mapped object at address 0x7073bde72000)