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 [ -n "$CLIENTS" ] || { skip "$0 Need two or more remote clients" && exit 0; }
29 [ $CLIENTCOUNT -ge 3 ] || \
30 { skip "$0 Need two or more remote clients, have $CLIENTCOUNT" && exit 0; }
32 END_RUN_FILE=${END_RUN_FILE:-$SHARED_DIRECTORY/end_run_file}
33 LOAD_PID_FILE=${LOAD_PID_FILE:-$TMP/client-load.pid}
35 remote_mds_nodsh && skip "remote MDS with nodsh" && exit 0
36 remote_ost_nodsh && skip "remote OST with nodsh" && exit 0
38 check_timeout || exit 1
42 check_and_setup_lustre
43 rm -rf $DIR/[df][0-9]*
45 # the test node needs to be insulated from a lustre failure as much as possible,
46 # so not even loading the lustre modules is ideal.
48 # -- remove hostname from clients list
49 zconf_umount $(hostname) $MOUNT
50 NODES_TO_USE=${NODES_TO_USE:-$CLIENTS}
51 NODES_TO_USE=$(exclude_items_from_list $NODES_TO_USE $(hostname))
53 check_progs_installed $NODES_TO_USE ${CLIENT_LOADS[@]}
55 MDTS=$(get_facets MDS)
56 OSTS=$(get_facets OST)
60 reboot_recover_node () {
61 # item var contains a pair of clients if nodetype=clients
62 # I would prefer to have a list here
65 local timeout=$($LCTL get_param -n timeout)
67 # MDS, OST item contains the facet
69 MDS|OST ) facet_failover $item
70 [ "$SERIAL" ] && wait_recovery_complete $item $((timeout * 4)) || true
72 clients) for c in ${item//,/ }; do
75 echo "Reintegrating $c"
76 zconf_mount $c $MOUNT || return $?
78 start_client_loads $item || return $?
80 * ) error "reboot_recover_node: nodetype=$nodetype. Must be one of 'MDS', 'OST', or 'clients'."
87 local excluded=${2:-""}
93 clients) list=$NODES_TO_USE
95 * ) error "Invalid type=$type. Must be one of 'MDS', 'OST', or 'clients'."
99 [ "$excluded" ] && list=$(exclude_items_from_list $list $excluded)
101 if [ ! "$(echo $list)" ]; then
106 item=$(get_random_entry $list)
107 if [ "$type" = clients ] ; then
108 item="$item $(get_random_entry $(exclude_items_from_list $list $item))"
109 item=$(comma_list $item)
116 # for the two nodetypes specified, chooses a random node(s) from each
117 # class, reboots the nodes sequentially, and then restarts lustre on
124 local client_nodes=""
131 ==== START === $title "
133 item1=$(get_item_type $type1)
135 { echo "type1=$type1 item1 is empty" && return 0; }
136 item2=$(get_item_type $type2 $item1)
138 { echo "type1=$type1 item1=$item1 type2=$type2 item2=$item2 is empty" && return 0; }
140 # Check that our client loads are still running. If any have died,
141 # that means they have died outside of recovery, which is unacceptable.
142 log "==== Checking the clients loads BEFORE failover -- failure NOT OK"
144 # FIXME. need print summary on exit
145 if ! check_client_loads $NODES_TO_USE; then
149 log "Done checking client loads. Failing type1=$type1 item1=$item1 ... "
151 reboot_recover_node $item1 $type1 || return $?
153 # Hendrix test17 description:
154 # Introduce a failure, wait at
155 # least 5 minutes (for recovery),
157 # failure, and wait another 5
160 # reboot_recover_node waits recovery in according to
162 # We have a "double failures" if SERIAL is not set,
163 # do not need a sleep between failures for "double failures"
165 log " Failing type2=$type2 item2=$item2 ... "
166 reboot_recover_node $item2 $type2 || return $?
168 # Client loads are allowed to die while in recovery, so we just
170 log "==== Checking the clients loads AFTER failovers -- ERRORS_OK=$ERRORS_OK"
171 restart_client_loads $NODES_TO_USE $ERRORS_OK || return $?
172 log "Done checking / re-Starting client loads. PASS"
176 summary_and_cleanup () {
180 # Having not empty END_RUN_FILE means the failed loads only
181 if [ -s $END_RUN_FILE ]; then
182 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
185 read END_RUN_NODE < $END_RUN_FILE
187 # a client load will end (i.e. fail) if it finds
188 # the end run file. that does not mean that that client load
189 # actually failed though. the first node in the END_RUN_NODE is
190 # the one we are really interested in.
191 if [ -n "$END_RUN_NODE" ]; then
192 var=${END_RUN_NODE}_load
193 echo "Client load failed on node $END_RUN_NODE"
195 echo "client $END_RUN_NODE load debug output :"
196 local logfile=${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}.debug
197 do_node ${END_RUN_NODE} "set -x; [ -e $logfile ] && cat $logfile " || true
202 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
203 echo "$0" >> $END_RUN_FILE
205 [ $rc -eq 0 ] || result=FAIL
208 Server failover period: $FAILOVER_PERIOD seconds
209 Exited after: $ELAPSED seconds
210 Status: $result: rc=$rc"
212 # make sure the client loads die
213 do_nodes $NODES_TO_USE "set -x; test -f $TMP/client-load.pid && \
214 { kill -s TERM \$(cat $TMP/client-load.pid) || true; }"
216 # and free up the pdshes that started them, if any are still around
217 if [ -n "$CLIENT_LOAD_PIDS" ]; then
218 kill $CLIENT_LOAD_PIDS || true
220 kill -9 $CLIENT_LOAD_PIDS || true
222 [ $rc -eq 0 ] && zconf_mount $(hostname) $MOUNT
226 trap summary_and_cleanup EXIT TERM INT
231 log "-----============= $0 starting =============-----"
237 # Set SERIAL to serialize the failure through a recovery of the first failure.
241 [ "$SERIAL" ] && ERRORS_OK=""
243 FAILOVER_PERIOD=${FAILOVER_PERIOD:-$((60*5))} # 5 minutes
245 # Start client loads.
246 start_client_loads $NODES_TO_USE
247 echo clients load pids:
248 if ! do_nodes $NODES_TO_USE "set -x; echo \$(hostname): && cat $TMP/client-load.pid"; then
249 if [ -e $DEBUGLOG ]; then
256 # FIXME: Do we want to have an initial sleep period where the clients
257 # just run before introducing a failure?
258 sleep $FAILOVER_PERIOD
261 failover_pair MDS OST "test 1: failover MDS, then OST =========="
262 sleep $FAILOVER_PERIOD
265 failover_pair MDS clients "test 2: failover MDS, then 2 clients ===="
266 sleep $FAILOVER_PERIOD
269 if [ $MDSCOUNT -gt 1 ]; then
270 failover_pair MDS MDS "test 3: failover MDS, then another MDS =="
271 sleep $FAILOVER_PERIOD
273 skip "$0 : $MDSCOUNT < 2 MDTs, test 3 skipped"
277 if [ $OSTCOUNT -gt 1 ]; then
278 failover_pair OST OST "test 4: failover OST, then another OST =="
279 sleep $FAILOVER_PERIOD
281 skip "$0 : $OSTCOUNT < 2 OSTs, test 4 skipped"
285 failover_pair OST clients "test 5: failover OST, then 2 clients ===="
286 sleep $FAILOVER_PERIOD
289 failover_pair OST MDS "test 6: failover OST, then MDS =========="
290 sleep $FAILOVER_PERIOD
293 failover_pair clients MDS "test 7: failover 2 clients, then MDS ===="
294 sleep $FAILOVER_PERIOD
297 #failover_pair clients OST "test 8: failover 2 clients, then OST ===="
298 sleep $FAILOVER_PERIOD
301 if [ $CLIENTCOUNT -ge 5 ]; then
302 failover_pair clients clients "test 9: failover 2 clients, then 2 different clients =="
303 sleep $FAILOVER_PERIOD
305 log "==== Checking the clients loads AFTER all failovers -- failure NOT OK"
306 if ! check_client_loads $NODES_TO_USE; then
307 log "Client load failed after failover. Exiting"
311 CURRENT_TS=$(date +%s)
312 ELAPSED=$((CURRENT_TS - START_TS))
314 log "Completed successfully in $ELAPSED seconds"