logos-storage-local-harness/test/test_process_monitor.bats

99 lines
1.9 KiB
Bash

setup() {
bats_require_minimum_version 1.12.0
export LIB_SRC="${BATS_TEST_DIRNAME}/../src"
load "${BATS_TEST_DIRNAME}/test_helper/bats-support/load.bash"
load "${BATS_TEST_DIRNAME}/test_helper/bats-assert/load.bash"
source "${LIB_SRC}/process_monitor.bash"
}
@test "should not start process monitor twice" {
assert_equal $(clh_monitor_state) "halted"
assert clh_start_process_monitor
assert_equal $(clh_monitor_state) "running"
refute clh_start_process_monitor
assert clh_stop_process_monitor
assert_equal $(clh_monitor_state) "halted"
}
@test "should not stop the process monitor if it wasn't started" {
refute clh_stop_process_monitor
}
@test "should keep track of process IDs" {
assert clh_start_process_monitor
clh_get_tracked_pids
assert [ ${#result[@]} -eq 0 ]
(
while [ ! -f "${_procmon_output}/sync" ]; do
sleep 0.1
done
clh_exit 0
) &
clh_track_last_background_job
p1=$!
(
while [ ! -f "${_procmon_output}/sync" ]; do
sleep 0.1
done
clh_exit 0
) &
clh_track_last_background_job
p2=$!
clh_get_tracked_pids
assert [ ${#result[@]} -eq 2 ]
touch "${_procmon_output}/sync"
await "$p1"
await "$p2"
# This should be more than enough for the process monitor to
# catch the exits. The alternative would be implementing temporal
# predicates.
sleep 1
clh_get_tracked_pids
assert [ ${#result[@]} -eq 0 ]
clh_stop_process_monitor
}
@test "should stop the monitor and all other processes if one process fails" {
assert clh_start_process_monitor
(
while [ ! -f "${_procmon_output}/sync" ]; do
sleep 0.1
done
clh_exit 1
) &
clh_track_last_background_job
p1=$!
(
while [ ! -f "${_procmon_output}/sync" ]; do
sleep 1
done
clh_exit 0
) &
clh_track_last_background_job
p2=$!
touch "${_procmon_output}/sync"
await "$p1"
await "$p2"
sleep 1
assert_equal $(clh_monitor_state) "halted_process_failure"
}