echo "$0: $(date +%H:%M:%S' '%s):" "$@"
}
+ha_touch()
+{
+ local date=$(date +%H:%M:%S' '%s)
+
+ [[ $1 =~ stop ]] &&
+ echo $date ${FUNCNAME[1]} $2 >> $ha_stop_file ||
+ true
+ [[ $1 =~ fail ]] &&
+ echo $date ${FUNCNAME[1]} $2 >> $ha_fail_file ||
+ true
+ [[ $1 =~ lfsck ]] &&
+ echo $date ${FUNCNAME[1]} $2 >> $ha_lfsck_stop ||
+ true
+}
+
ha_log()
{
local nodes=${1// /,}
trap ha_trap_err ERR
set -eE
+declare ha_power_down_pids
declare ha_tmp_dir=/tmp/$(basename $0)-$$
declare ha_stop_file=$ha_tmp_dir/stop
declare ha_fail_file=$ha_tmp_dir/fail
declare ha_power_up_cmd=${POWER_UP:-"pm -1"}
declare ha_power_delay=${POWER_DELAY:-60}
declare ha_node_up_delay=${NODE_UP_DELAY:-10}
+declare ha_wait_nodes_up=${WAIT_NODES_UP:-600}
declare ha_pm_host=${PM_HOST:-$(hostname)}
declare ha_failback_delay=${DELAY:-5}
declare ha_failback_cmd=${FAILBACK:-""}
declare ha_test_dir_mdt_index=${TDMDTINDEX:-"0"}
declare ha_test_dir_mdt_index_random=${TDMDTINDEXRAND:-false}
declare ha_dir_stripe_count=${DSTRIPECOUNT:-"1"}
+declare ha_dir_stripe_count_random=${DSTRIPECOUNTRAND:-false}
declare ha_mdt_index=${MDTINDEX:-"0"}
declare ha_mdt_index_random=${MDTINDEXRAND:-false}
declare -a ha_clients
declare -a ha_victims
declare -a ha_victims_pair
declare ha_test_dir=/mnt/lustre/$(basename $0)-$$
+declare -a ha_testdirs=(${ha_test_dirs="$ha_test_dir"})
+
+for ((i=0; i<${#ha_testdirs[@]}; i++)); do
+ echo I=$i ${ha_testdirs[i]}
+ ha_testdirs[i]="${ha_testdirs[i]}/$(basename $0)-$$"
+ echo i=$i ${ha_testdirs[i]}
+done
+
+declare ha_cleanup=${CLEANUP:-true}
declare ha_start_time=$(date +%s)
declare ha_expected_duration=$((60 * 60 * 24))
declare ha_max_failover_period=10
[iozone]="iozone -a -e -+d -s $iozone_SIZE {}/f.iozone"
[racer]="$ha_racer_params $racer {}"
)
+declare ha_check_attrs="find {} -type f -ls 2>&1 | grep -e '?'"
ha_usage()
{
ha_trap_exit()
{
- touch "$ha_stop_file"
+ ha_touch stop
trap 0
if [ -e "$ha_fail_file" ]; then
- ha_info "Test directory $ha_test_dir not removed"
+ ha_info "Test directories ${ha_testdirs[@]} not removed"
ha_info "Temporary directory $ha_tmp_dir not removed"
else
- ha_on ${ha_clients[0]} rm -rf "$ha_test_dir"
+ $ha_cleanup &&
+ ha_on ${ha_clients[0]} rm -rf ${ha_testdirs[@]} ||
+ ha_info "Test directories ${ha_testdirs[@]} not removed"
ha_info "Please find the results in the directory $ha_tmp_dir"
fi
}
ha_trap_stop_signals()
{
- ha_info "${ha_stop_signals// /,} received"
- touch "$ha_stop_file"
+ ha_info "${ha_stop_signals// /,} received"
+ ha_touch stop "${ha_stop_signals// /,} received"
}
ha_sleep()
local mpiuser=$7
local mustpass=$8
local mpirunoptions=$9
+ local test_dir=${10}
local tag=${ha_mpi_load_tags[$load]}
local cmd=${ha_mpi_load_cmds[$tag]}
- local dir=$ha_test_dir/$client-$tag
+ local dir=$test_dir/$client-$tag
local log=$ha_tmp_dir/$client-$tag
local rc=0
+ local rccheck=0
local nr_loops=0
local avg_loop_time=0
local start_time=$(date +%s)
+ local check_attrs=${ha_check_attrs//"{}"/$dir}
cmd=${cmd//"{}"/$dir}
cmd=${cmd//"{params}"/$parameter}
ha_info "Starting $tag"
machines="-machinefile $machines"
- while [ ! -e "$ha_stop_file" ] && ((rc == 0)); do
+ while [ ! -e "$ha_stop_file" ] && ((rc == 0)) && ((rccheck == 0)); do
ha_info "$client Starts: $mpiuser: $cmd" 2>&1 | tee -a $log
{
local mdt_index
if $ha_mdt_index_random && [ $ha_mdt_index -ne 0 ]; then
- mdt_index=$(ha_rand $ha_mdt_index)
+ mdt_index=$(ha_rand $((ha_mdt_index + 1)) )
else
mdt_index=$ha_mdt_index
fi
+ local dir_stripe_count
+ if $ha_dir_stripe_count_random &&
+ [ $ha_dir_stripe_count -ne 1 ]; then
+ dir_stripe_count=$(($(ha_rand $ha_dir_stripe_count) + 1))
+ else
+ dir_stripe_count=$ha_dir_stripe_count
+ fi
[[ -n "$ha_precmd" ]] && ha_info "$ha_precmd" &&
ha_on $client "$ha_precmd" >>"$log" 2>&1
- ha_on $client $LFS mkdir -i$mdt_index -c$ha_dir_stripe_count "$dir" &&
+ ha_info "$client Creates $dir with -i$mdt_index -c$dir_stripe_count "
+ ha_on $client $LFS mkdir -i$mdt_index -c$dir_stripe_count "$dir" &&
ha_on $client $LFS getdirstripe "$dir" &&
ha_on $client $LFS setstripe $stripeparams $dir &&
ha_on $client $LFS getstripe $dir &&
ha_on $client chmod a+xwr $dir &&
- ha_on $client "su $mpiuser sh -c \" $mpirun $mpirunoptions \
- -np $((${#ha_clients[@]} * mpi_threads_per_client )) \
+ ha_on $client "su $mpiuser bash -c \" $mpirun $mpirunoptions \
+ -np $((${#ha_clients[@]} * mpi_threads_per_client / ha_nclientsset)) \
$machines $cmd \" " || rc=$?
+ ha_on ${ha_clients[0]} "$check_attrs && \
+ $LFS df $dir && \
+ $check_attrs " && rccheck=1
[[ -n "$ha_postcmd" ]] && ha_info "$ha_postcmd" &&
ha_on $client "$ha_postcmd" >>"$log" 2>&1
- (( ((rc == 0)) && (( mustpass != 0 )) )) ||
- (( ((rc != 0)) && (( mustpass == 0 )) )) &&
- ha_on $client rm -rf "$dir";
- } >>"$log" 2>&1 || rc=$?
+ (( ((rc == 0)) && ((rccheck == 0)) && (( mustpass != 0 )) )) ||
+ (( ((rc != 0)) && ((rccheck == 0)) && (( mustpass == 0 )) )) &&
+ ha_on $client rm -rf "$dir";
+ } >>"$log" 2>&1
- ha_info $client: rc=$rc mustpass=$mustpass
+ ha_info $client: rc=$rc rccheck=$rccheck mustpass=$mustpass
# mustpass=0 means that failure is expected
- if (( rc !=0 )); then
+ if (( rccheck != 0 )); then
+ ha_touch stop,fail $client,$tag
+ ha_dump_logs "${ha_clients[*]} ${ha_servers[*]}"
+ elif (( rc !=0 )); then
if (( mustpass != 0 )); then
- touch "$ha_fail_file"
- touch "$ha_stop_file"
+ ha_touch stop,fail $client,$tag
ha_dump_logs "${ha_clients[*]} ${ha_servers[*]}"
else
# Ok to fail
rc=0
fi
elif (( mustpass == 0 )); then
- touch "$ha_fail_file"
- touch "$ha_stop_file"
+ ha_touch stop,fail $client,$tag
ha_dump_logs "${ha_clients[*]} ${ha_servers[*]}"
fi
- echo rc=$rc mustpass=$mustpass >"$status"
+ echo rc=$rc rccheck=$rccheck mustpass=$mustpass >"$status"
nr_loops=$((nr_loops + 1))
done
scp $ha_machine_file* $client:$dirname
done
+ local ndir
for ((n = 0; n < $inst; n++)); do
client=${ha_clients[n]}
nmpi=$((n % ${#ha_users[@]}))
mpiuser=${ha_users[nmpi]}
+ ndir=$((n % ${#ha_testdirs[@]}))
+ test_dir=${ha_testdirs[ndir]}
for ((load = 0; load < ${#ha_mpi_load_tags[@]}; load++)); do
tag=${ha_mpi_load_tags[$load]}
status=$ha_status_file_prefix-$tag-$client
mustpass=$(( n % ha_ninstmustfail ))
ha_repeat_mpi_load $client $load $status "$parameter" \
$machines "$stripe" "$mpiuser" "$mustpass" \
- "${ha_mpiopts[$mpiuser]} $ha_mpirun_options" &
+ "${ha_mpiopts[$mpiuser]} $ha_mpirun_options" "$test_dir" &
ha_status_files+=("$status")
done
done
local status=$3
local tag=${ha_nonmpi_load_tags[$load]}
local cmd=${ha_nonmpi_load_cmds[$tag]}
- local dir=$ha_test_dir/$client-$tag
+ local test_dir=$4
+ local dir=$test_dir/$client-$tag
local log=$ha_tmp_dir/$client-$tag
local rc=0
+
+ local rccheck=0
local nr_loops=0
local avg_loop_time=0
local start_time=$(date +%s)
+ local check_attrs=${ha_check_attrs//"{}"/$dir}
cmd=${cmd//"{}"/$dir}
- ha_info "Starting $tag on $client"
+ ha_info "Starting $tag on $client on $dir"
while [ ! -e "$ha_stop_file" ] && ((rc == 0)); do
ha_info "$client Starts: $cmd" 2>&1 | tee -a $log
ha_on $client "mkdir -p $dir && \
- $cmd && \
- rm -rf $dir" >>"$log" 2>&1 || rc=$?
+ $cmd" >>"$log" 2>&1 || rc=$?
+
+ ha_on $client "$check_attrs && \
+ $LFS df $dir && \
+ $check_attrs " >>"$log" 2>&1 && rccheck=1 ||
+ ha_on $client "rm -rf $dir" >>"$log" 2>&1
- if ((rc != 0)); then
+ ha_info rc=$rc rccheck=$rccheck
+
+ if (( (rc + rccheck) != 0 )); then
ha_dump_logs "${ha_clients[*]} ${ha_servers[*]}"
- touch "$ha_fail_file"
- touch "$ha_stop_file"
+ ha_touch stop,fail $client,$tag
fi
echo $rc >"$status"
ha_start_nonmpi_loads()
{
- local client
- local load
- local tag
- local status
-
- for client in ${ha_clients[@]}; do
- for ((load = 0; load < ${#ha_nonmpi_load_tags[@]}; load++)); do
- tag=${ha_nonmpi_load_tags[$load]}
- status=$ha_status_file_prefix-$tag-$client
- ha_repeat_nonmpi_load $client $load $status &
- ha_status_files+=("$status")
- done
- done
+ local client
+ local load
+ local tag
+ local status
+ local n
+ local test_dir
+ local ndir
+
+ for (( n = 0; n < ${#ha_clients[@]}; n++)); do
+ client=${ha_clients[n]}
+ ndir=$((n % ${#ha_testdirs[@]}))
+ test_dir=${ha_testdirs[ndir]}
+ for ((load = 0; load < ${#ha_nonmpi_load_tags[@]}; load++)); do
+ tag=${ha_nonmpi_load_tags[$load]}
+ status=$ha_status_file_prefix-$tag-$client
+ ha_repeat_nonmpi_load $client $load $status $test_dir &
+ ha_status_files+=("$status")
+ done
+ done
+}
+
+declare ha_bgcmd=${ha_bgcmd:-""}
+declare ha_bgcmd_log=$ha_tmp_dir/bgcmdlog
+
+ha_cmd_bg () {
+ [[ -z "$ha_bgcmd" ]] && return 0
+ for ((i=0; i<${#ha_testdirs[@]}; i++)); do
+ ha_bgcmd=${ha_bgcmd//"{}"/${ha_testdirs[i]}}
+ done
+
+ ha_info "BG cmd: $ha_bgcmd"
+ while [ true ]; do
+ [ -f $ha_stop_file ] &&
+ ha_info "$ha_stop_file found! $ha_bgcmd no started" &&
+ break
+ eval $ha_bgcmd 2>&1 | tee -a $ha_bgcmd_log
+ sleep 1
+ done &
+ CMD_BG_PID=$!
+ ha_info CMD BG PID: $CMD_BG_PID
+ ps aux | grep $CMD_BG_PID
}
ha_lfsck_bg () {
fi
[ $rc -eq 0 ] ||
- { touch "$ha_fail_file"; touch "$ha_stop_file";
- touch $ha_lfsck_stop; return 1; }
+ { ha_touch stop,fail,lfsck; return 1; }
ha_wait_lfsck_completed ||
- { touch "$ha_fail_file"; touch "$ha_stop_file";
- touch $ha_lfsck_stop; return 1; }
+ { ha_touch stop,fail,lfsck; return 1; }
return 0
}
awk '{sum += $1} END { print sum }')
[ $n -eq 0] ||
{ ha_info "Total repaired: $n";
- touch "$ha_fail_file"; return 1; }
+ ha_touch fail; return 1; }
return 0
}
ha_start_loads()
{
+ ha_cmd_bg
$ha_lfsck_bg && ha_lfsck_bg
trap ha_trap_stop_signals $ha_stop_signals
ha_start_nonmpi_loads
ha_stop_loads()
{
- touch $ha_stop_file
+ ha_touch stop
+ [[ -n $CMD_BG_PID ]] && wait $CMD_BG_PID || true
# true because of lfsck_bg could be stopped already
$ha_lfsck_bg && wait $LFSCK_BG_PID || true
trap - $ha_stop_signals
{
local nodes=$1
local cmd
+ local pid
+ local rc=0
case $ha_power_down_cmd in
# format is: POWER_DOWN=sysrqcrash
- sysrqcrash) cmd="pdsh -S -w $nodes 'echo c > /proc/sysrq-trigger' &" ;;
- *) cmd="$ha_power_down_cmd $nodes" ;;
+ sysrqcrash)
+ cmd="pdsh -S -w $nodes -u 120 \"echo c > /proc/sysrq-trigger\" &"
+ eval $cmd
+ pid=$!
+ ha_power_down_pids=$(echo $ha_power_down_pids $pid)
+ ha_info "ha_power_down_pids: $ha_power_down_pids"
+ [[ -z "$ha_power_down_pids" ]] ||
+ ps aux | grep " ${ha_power_down_pids// / \| } " ||
+ true
+ ;;
+ *)
+ cmd="$ha_power_down_cmd $nodes"
+ eval $cmd
+ rc=$?
+ ;;
esac
- eval $cmd
+ return $rc
}
ha_power_down()
fi
ha_info "Powering down $nodes : cmd: $ha_power_down_cmd"
+ ha_power_down_pids=""
for (( i=0; i<10; i++ )) {
ha_info "attempt: $i"
- ha_power_down_cmd_fn $nodes &&
- ha_powermanage $nodes $state && rc=0 && break
- sleep $ha_power_delay
+ ha_power_down_cmd_fn $nodes || rc=1
+ ha_sleep $ha_power_delay
+ ha_powermanage $nodes $state && rc=0 && break
}
+ if [[ -n "$ha_power_down_pids" ]]; then
+ kill -9 $ha_power_down_pids || true
+ wait $ha_power_down_pids || true
+ fi
[ $rc -eq 0 ] || {
ha_info "Failed Powering down in $i attempts:" \
ha_wait_nodes()
{
local nodes=$1
- local end=$(($(date +%s) + 10 * 60))
+ local end=$(($(date +%s) + $ha_wait_nodes_up))
- ha_info "Waiting for $nodes to boot up"
+ ha_info "Waiting for $nodes to boot up in $ha_wait_nodes_up"
until ha_on $nodes hostname >/dev/null 2>&1 ||
[ -e "$ha_stop_file" ] ||
(($(date +%s) >= end)); do
ha_sleep 1 >/dev/null
done
+
+ ha_info "Check where we are ..."
+ [ -e "$ha_stop_file" ] &&
+ ha_info "$ha_stop_file found!"
+
+ local -a nodes_up
+ nodes_up=($(ha_on $nodes hostname | awk '{ print $2 }'))
+ ha_info "Nodes $nodes are up: ${nodes_up[@]}"
+ local -a n=(${nodes//,/ })
+ if [[ ${#nodes_up[@]} -ne ${#n[@]} ]]; then
+ ha_info "Failed boot up $nodes in $ha_wait_nodes_up sec!"
+ ha_touch fail,stop
+ return 1
+ fi
+ return 0
}
ha_failback()
local mdt_index
if $ha_test_dir_mdt_index_random &&
[ $ha_test_dir_mdt_index -ne 0 ]; then
- mdt_index=$(ha_rand $ha_test_dir_mdt_index)
+ mdt_index=$(ha_rand $((ha_test_dir_mdt_index + 1)) )
else
mdt_index=$ha_test_dir_mdt_index
fi
+
+ local dir
+ test_dir=${ha_testdirs[0]}
ha_on ${ha_clients[0]} "$LFS mkdir -i$mdt_index \
- -c$ha_test_dir_stripe_count $ha_test_dir"
- ha_on ${ha_clients[0]} $LFS getdirstripe $ha_test_dir
- ha_on ${ha_clients[0]} " \
- $LFS setstripe $ha_stripe_params $ha_test_dir"
+ -c$ha_test_dir_stripe_count $test_dir"
+ for ((i=0; i<${#ha_testdirs[@]}; i++)); do
+ test_dir=${ha_testdirs[i]}
+ ha_on ${ha_clients[0]} $LFS getdirstripe $test_dir
+ ha_on ${ha_clients[0]} " \
+ $LFS setstripe $ha_stripe_params $test_dir"
+ done
ha_start_loads
ha_wait_loads