fence upcall not issued for rank subset

[garlick]

This test, which attempts a PMIx_Fence() with ranks set to a subset of the full job, currently hangs. The fence upcall is never received (it blocks in openpmix). What we would like to do is fail it, since currently the implementation does not support fence over a subset.

expecting success: 
	test_must_fail run_timeout 30 flux mini run -N2 -n2 \
		-ouserrc=$(pwd)/rc.lua \
		-overbose=2 \
		${BARRIER} --procs=1

0.022s: flux-shell[1]: DEBUG: 1: tasks [1] on cores 0
0.022s: flux-shell[1]: DEBUG: Loading /opt/flux-core-0.29.0-27/etc/flux/shell/initrc.lua
0.041s: flux-shell[1]: TRACE: shell barrier complete
0.041s: flux-shell[1]: TRACE: exited barrier with rc = 0
0.044s: flux-shell[1]: TRACE: shell barrier complete
0.044s: flux-shell[1]: TRACE: exited barrier with rc = 0
0.019s: flux-shell[0]: DEBUG: 0: task_count=2 slot_count=2 cores_per_slot=1 slots_per_node=1
0.019s: flux-shell[0]: DEBUG: 0: tasks [0] on cores 0
0.020s: flux-shell[0]: DEBUG: Loading /opt/flux-core-0.29.0-27/etc/flux/shell/initrc.lua
0.022s: flux-shell[0]: DEBUG: output: batch timeout = 0.500s
0.041s: flux-shell[0]: TRACE: shell barrier complete
0.041s: flux-shell[0]: TRACE: exited barrier with rc = 0
0.044s: flux-shell[0]: TRACE: shell barrier complete
0.044s: flux-shell[0]: TRACE: exited barrier with rc = 0
0.053s: flux-shell[1]: TRACE: 1: C: pmi EOF
0.053s: flux-shell[1]: DEBUG: task 1 complete status=0
0.064s: flux-shell[1]: DEBUG: exit 0
f24XVCK1.0: completed PMIx_Init.
f24XVCK1.0: there are 2 tasks
test_must_fail: died by signal: run_timeout 30 flux mini run -N2 -n2 -ouserrc=/home/garlick/proj/flux-pmix/t/trash-directory.t0003-barrier/rc.lua -overbose=2 /home/garlick/proj/flux-pmix/t/src/barrier --procs=1
not ok 15 - 2n2p barrier with procs subset fails

[garlick]

Note that the proc on shell 1 (task 1) exits early. So I guess the fence call returned immediately because that proc is not participating. Meanwhile the other proc is likely blocked in the fence call.

It would seem like this should look like a degenerate case to the pmix library (fence call only involving self) and return immedaitely too, but it does not.

[garlick]

Also, just noticed that the "baseline" test using prterun with a rank subset intermittently fails

test_expect_success 'prterun barrier works with procs subset' '
        ${PRTERUN} --n 8 ${BARRIER} --procs=0-1
'

Mark it XFAIL and just avoid this case that is not likely found in nature anyway.