3 # All pairwise combinations of node failures.
6 # Author: Chris Cooper <ccooper@clusterfs.com>
8 # Script fails pair of nodes:
9 # -- in parallel by default
10 # -- in series if SERIAL is set
12 LUSTRE=${LUSTRE:-`dirname $0`/..}
14 CLEANUP=${CLEANUP:-""}
15 . $LUSTRE/tests/test-framework.sh
19 . ${CONFIG:=$LUSTRE/tests/cfg/$NAME.sh}
20 TESTSUITELOG=${TESTSUITELOG:-$TMP/recovery-double-scale}
21 DEBUGLOG=$TESTSUITELOG.debug
23 echo "--- env ---" >&2
25 echo "--- env ---" >&2
28 [ "$SHARED_DIRECTORY" ] || \
29 { skip "$0: Empty SHARED_DIRECTORY" && exit 0; }
31 [ -n "$CLIENTS" ] || { skip "$0 Need two or more remote clients" && exit 0; }
32 [ $CLIENTCOUNT -ge 3 ] || \
33 { skip "$0 Need two or more remote clients, have $CLIENTCOUNT" && exit 0; }
35 END_RUN_FILE=${END_RUN_FILE:-$SHARED_DIRECTORY/end_run_file}
36 LOAD_PID_FILE=${LOAD_PID_FILE:-$TMP/client-load.pid}
38 remote_mds_nodsh && skip "remote MDS with nodsh" && exit 0
39 remote_ost_nodsh && skip "remote OST with nodsh" && exit 0
41 check_timeout || exit 1
45 check_and_setup_lustre
46 rm -rf $DIR/[df][0-9]*
48 # the test node needs to be insulated from a lustre failure as much as possible,
49 # so not even loading the lustre modules is ideal.
51 # -- remove hostname from clients list
52 zconf_umount $(hostname) $MOUNT
53 NODES_TO_USE=${NODES_TO_USE:-$CLIENTS}
54 NODES_TO_USE=$(exclude_items_from_list $NODES_TO_USE $(hostname))
56 check_progs_installed $NODES_TO_USE ${CLIENT_LOADS[@]}
58 MDTS=$(get_facets MDS)
59 OSTS=$(get_facets OST)
63 reboot_recover_node () {
64 # item var contains a pair of clients if nodetype=clients
65 # I would prefer to have a list here
68 local timeout=$($LCTL get_param -n timeout)
70 # MDS, OST item contains the facet
72 MDS|OST ) facet_failover $item
73 [ "$SERIAL" ] && wait_recovery_complete $item $((timeout * 4)) || true
75 clients) for c in ${item//,/ }; do
78 echo "Reintegrating $c"
79 # one client fails; need dk logs from this client only
80 zconf_mount $c $MOUNT || NODES="$c $(mdts_nodes) $(osts_nodes)" error_exit "zconf_mount failed"
82 start_client_loads $item
85 # don't use error (), the logs from all nodes not needed
86 * ) echo "reboot_recover_node: nodetype=$nodetype. Must be one of 'MDS', 'OST', or 'clients'."
93 local excluded=${2:-""}
99 clients) list=$NODES_TO_USE
102 # don't use error (), the logs from all nodes not needed
103 * ) echo "Invalid type=$type. Must be one of 'MDS', 'OST', or 'clients'."
107 [ "$excluded" ] && list=$(exclude_items_from_list $list $excluded)
109 if [ ! "$(echo $list)" ]; then
114 item=$(get_random_entry $list)
115 if [ "$type" = clients ] ; then
116 item="$item $(get_random_entry $(exclude_items_from_list $list $item))"
117 item=$(comma_list $item)
124 # for the two nodetypes specified, chooses a random node(s) from each
125 # class, reboots the nodes sequentially, and then restarts lustre on
132 local client_nodes=""
139 ==== START === $title "
141 item1=$(get_item_type $type1)
143 { echo "type1=$type1 item1 is empty" && return 0; }
144 item2=$(get_item_type $type2 $item1)
146 { echo "type1=$type1 item1=$item1 type2=$type2 item2=$item2 is empty" && return 0; }
148 # Check that our client loads are still running. If any have died,
149 # that means they have died outside of recovery, which is unacceptable.
150 log "==== Checking the clients loads BEFORE failover -- failure NOT OK"
152 # FIXME. need print summary on exit
153 if ! check_client_loads $NODES_TO_USE; then
157 log "Done checking client loads. Failing type1=$type1 item1=$item1 ... "
159 reboot_recover_node $item1 $type1
161 # Hendrix test17 description:
162 # Introduce a failure, wait at
163 # least 5 minutes (for recovery),
165 # failure, and wait another 5
168 # reboot_recover_node waits recovery in according to
170 # We have a "double failures" if SERIAL is not set,
171 # do not need a sleep between failures for "double failures"
173 log " Failing type2=$type2 item2=$item2 ... "
174 reboot_recover_node $item2 $type2
176 # Client loads are allowed to die while in recovery, so we just
178 log "==== Checking the clients loads AFTER failovers -- ERRORS_OK=$ERRORS_OK"
179 restart_client_loads $NODES_TO_USE $ERRORS_OK || return $?
180 log "Done checking / re-Starting client loads. PASS"
184 summary_and_cleanup () {
188 # Having not empty END_RUN_FILE means the failed loads only
189 if [ -s $END_RUN_FILE ]; then
190 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
193 read END_RUN_NODE < $END_RUN_FILE
195 # a client load will end (i.e. fail) if it finds
196 # the end run file. that does not mean that that client load
197 # actually failed though. the first node in the END_RUN_NODE is
198 # the one we are really interested in.
199 if [ -n "$END_RUN_NODE" ]; then
200 var=${END_RUN_NODE}_load
201 echo "Client load failed on node $END_RUN_NODE"
203 echo "client $END_RUN_NODE load debug output :"
204 local logfile=${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}.debug
205 do_node ${END_RUN_NODE} "set -x; [ -e $logfile ] && cat $logfile " || true
210 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
211 echo "$0" >> $END_RUN_FILE
213 [ $rc -eq 0 ] || result=FAIL
216 Server failover period: $FAILOVER_PERIOD seconds
217 Exited after: $ELAPSED seconds
218 Status: $result: rc=$rc"
220 # make sure the client loads die
221 do_nodes $NODES_TO_USE "set -x; test -f $TMP/client-load.pid && \
222 { kill -s TERM \$(cat $TMP/client-load.pid) || true; }"
224 # and free up the pdshes that started them, if any are still around
225 if [ -n "$CLIENT_LOAD_PIDS" ]; then
226 kill $CLIENT_LOAD_PIDS || true
228 kill -9 $CLIENT_LOAD_PIDS || true
230 [ $rc -eq 0 ] && zconf_mount $(hostname) $MOUNT
234 trap summary_and_cleanup EXIT TERM INT
239 log "-----============= $0 starting =============-----"
245 # Set SERIAL to serialize the failure through a recovery of the first failure.
249 [ "$SERIAL" ] && ERRORS_OK=""
251 FAILOVER_PERIOD=${FAILOVER_PERIOD:-$((60*5))} # 5 minutes
253 # Start client loads.
254 start_client_loads $NODES_TO_USE
255 echo clients load pids:
256 if ! do_nodes $NODES_TO_USE "set -x; echo \$(hostname): && cat $TMP/client-load.pid"; then
257 if [ -e $DEBUGLOG ]; then
264 # FIXME: Do we want to have an initial sleep period where the clients
265 # just run before introducing a failure?
266 sleep $FAILOVER_PERIOD
269 failover_pair MDS OST "test 1: failover MDS, then OST =========="
270 sleep $FAILOVER_PERIOD
273 failover_pair MDS clients "test 2: failover MDS, then 2 clients ===="
274 sleep $FAILOVER_PERIOD
277 # No test 3 for 1.8.x lustre version
280 if [ $OSTCOUNT -gt 1 ]; then
281 failover_pair OST OST "test 4: failover OST, then another OST =="
282 sleep $FAILOVER_PERIOD
284 skip "$0 : $OSTCOUNT < 2 OSTs, test 4 skipped"
288 failover_pair OST clients "test 5: failover OST, then 2 clients ===="
289 sleep $FAILOVER_PERIOD
292 failover_pair OST MDS "test 6: failover OST, then MDS =========="
293 sleep $FAILOVER_PERIOD
296 failover_pair clients MDS "test 7: failover 2 clients, then MDS ===="
297 sleep $FAILOVER_PERIOD
300 #failover_pair clients OST "test 8: failover 2 clients, then OST ===="
301 sleep $FAILOVER_PERIOD
304 if [ $CLIENTCOUNT -ge 5 ]; then
305 failover_pair clients clients "test 9: failover 2 clients, then 2 different clients =="
306 sleep $FAILOVER_PERIOD
308 log "==== Checking the clients loads AFTER all failovers -- failure NOT OK"
309 if ! check_client_loads $NODES_TO_USE; then
310 log "Client load failed after failover. Exiting"
314 CURRENT_TS=$(date +%s)
315 ELAPSED=$((CURRENT_TS - START_TS))
317 log "Completed successfully in $ELAPSED seconds"