4 # For duration of 24 hours repeatedly failover a random MDS at
5 # 10 minute intervals and verify that no application errors occur.
7 # Test runs one of CLIENT_LOAD progs on remote clients.
9 LUSTRE=${LUSTRE:-`dirname $0`/..}
11 CLEANUP=${CLEANUP:-""}
12 . $LUSTRE/tests/test-framework.sh
16 . ${CONFIG:=$LUSTRE/tests/cfg/$NAME.sh}
19 TESTSUITELOG=${TESTSUITELOG:-$TMP/$(basename $0 .sh)}
20 DEBUGLOG=$TESTSUITELOG.debug
25 echo "--- env ---" >&2
27 echo "--- env ---" >&2
30 [ "$SHARED_DIRECTORY" ] || \
31 { skip "$0: Empty SHARED_DIRECTORY" && exit 0; }
33 [ -n "$CLIENTS" ] || { skip "$0 Need two or more remote clients" && exit 0; }
34 [ $CLIENTCOUNT -ge 3 ] || \
35 { skip "$0 Need two or more remote clients, have $CLIENTCOUNT" && exit 0; }
37 END_RUN_FILE=${END_RUN_FILE:-$SHARED_DIRECTORY/end_run_file}
38 LOAD_PID_FILE=${LOAD_PID_FILE:-$TMP/client-load.pid}
40 remote_mds_nodsh && skip "remote MDS with nodsh" && exit 0
41 remote_ost_nodsh && skip "remote OST with nodsh" && exit 0
45 check_and_setup_lustre
46 rm -rf $DIR/[df][0-9]*
48 max_recov_time=$(max_recovery_time)
50 # the test node needs to be insulated from a lustre failure as much as possible,
51 # so not even loading the lustre modules is ideal.
53 # -- remove hostname from clients list
54 zconf_umount $(hostname) $MOUNT
55 NODES_TO_USE=${NODES_TO_USE:-$CLIENTS}
56 NODES_TO_USE=$(exclude_items_from_list $NODES_TO_USE $(hostname))
58 check_progs_installed $NODES_TO_USE ${CLIENT_LOADS[@]}
60 MDTS=$(get_facets MDS)
61 OSTS=$(get_facets OST)
63 ERRORS_OK="" # No application failures should occur during this test.
64 FLAVOR=${FLAVOR:-"MDS"}
66 if [ "$FLAVOR" == "MDS" ]; then
72 if [ "$SLOW" = "no" ]; then
73 DURATION=${DURATION:-$((60 * 30))}
74 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 5))}
76 DURATION=${DURATION:-$((60 * 60 * 24))}
77 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 10))} # 10 minutes
82 vmstatLOG=${TESTSUITELOG}_$(basename $0 .sh).vmstat
84 server_numfailovers () {
86 local var=${facet}_numfailovers
89 [[ ${!var} ]] && val=${!var}
93 servers_numfailovers () {
97 for facet in ${MDTS//,/ } ${OSTS//,/ }; do
98 echo "$facet: $(server_numfailovers $facet) times"
102 summary_and_cleanup () {
108 # Having not empty END_RUN_FILE means the failed loads only
109 if [ -s $END_RUN_FILE ]; then
110 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
113 read END_RUN_NODE < $END_RUN_FILE
115 # a client load will end (i.e. fail) if it finds
116 # the end run file. that does not mean that that client load
117 # actually failed though. the first node in the END_RUN_NODE is
118 # the one we are really interested in.
119 if [ -n "$END_RUN_NODE" ]; then
120 var=$(client_var_name $END_RUN_NODE)_load
121 echo "Client load failed on node $END_RUN_NODE"
123 echo "client $END_RUN_NODE load stdout and debug files :
124 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}
125 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}.debug"
130 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
131 echo "$0" >> $END_RUN_FILE
133 [ $rc -eq 0 ] || result=FAIL
135 log "Duration: $DURATION
136 Server failover period: $SERVER_FAILOVER_PERIOD seconds
137 Exited after: $ELAPSED seconds
138 Number of failovers before exit:
139 $(servers_numfailovers)
140 Status: $result: rc=$rc"
142 # stop the vmstats on the OSTs
143 if [ "$VMSTAT" ]; then
144 do_nodes $(comma_list $(osts_nodes)) "test -f /tmp/vmstat.pid && \
145 { kill -s TERM \$(cat /tmp/vmstat.pid); rm -f /tmp/vmstat.pid; \
146 gzip -f9 $vmstatLOG-\$(hostname); }"
149 # make sure the client loads die
150 do_nodes $NODES_TO_USE "set -x; test -f $LOAD_PID_FILE && \
151 { kill -s TERM \$(cat $LOAD_PID_FILE) || true; }"
153 # and free up the pdshes that started them, if any are still around
154 if [ -n "$CLIENT_LOAD_PIDS" ]; then
155 kill $CLIENT_LOAD_PIDS || true
157 kill -9 $CLIENT_LOAD_PIDS || true
159 if [ $rc -ne 0 ]; then
160 # we are interested in only on failed clients and servers
161 local failedclients=$(cat $END_RUN_FILE | grep -v $0)
162 # FIXME: need ostfailover-s nodes also for FLAVOR=OST
163 local product=$(gather_logs $(comma_list $(osts_nodes) \
164 $(mdts_nodes) $mdsfailover_HOST $failedclients))
165 echo logs files $product
168 [ $rc -eq 0 ] && zconf_mount $(hostname) $MOUNT
176 log "-----============= $0 starting =============-----"
178 trap summary_and_cleanup EXIT INT
183 if [ "$VMSTAT" ]; then
184 do_nodes $(comma_list $(osts_nodes)) "vmstat 1 > $vmstatLOG-\$(hostname) 2>/dev/null </dev/null & echo \$! > /tmp/vmstat.pid"
187 # Start client loads.
188 start_client_loads $NODES_TO_USE
190 echo clients load pids:
191 if ! do_nodesv $NODES_TO_USE "cat $LOAD_PID_FILE"; then
195 MINSLEEP=${MINSLEEP:-120}
196 REQFAIL_PERCENT=${REQFAIL_PERCENT:-3} # bug17839 comment 62
197 REQFAIL=${REQFAIL:-$(( DURATION / SERVER_FAILOVER_PERIOD * REQFAIL_PERCENT / 100))}
204 while [ $ELAPSED -lt $DURATION -a ! -e $END_RUN_FILE ]; do
206 # In order to perform the
207 # expected number of failovers, we need to account the following :
208 # 1) the time that has elapsed during the client load checking
209 # 2) time takes for failover
211 it_time_start=$(date +%s)
213 SERVERFACET=$(get_random_entry $SERVERS)
214 var=${SERVERFACET}_numfailovers
216 # Check that our client loads are still running. If any have died,
217 # that means they have died outside of recovery, which is unacceptable.
219 log "==== Checking the clients loads BEFORE failover -- failure NOT OK \
220 ELAPSED=$ELAPSED DURATION=$DURATION PERIOD=$SERVER_FAILOVER_PERIOD"
222 if ! check_client_loads $NODES_TO_USE; then
226 log "Wait $SERVERFACET recovery complete before doing next failover ...."
227 if [[ $(server_numfailovers $SERVERFACET) != 0 ]]; then
228 if ! wait_recovery_complete $SERVERFACET ; then
229 echo "$SERVERFACET recovery is not completed!"
234 log "Checking clients are in FULL state before doing next failover"
235 if ! wait_clients_import_state $NODES_TO_USE $SERVERFACET FULL; then
236 echo "Clients import not FULL, please consider to increase SERVER_FAILOVER_PERIOD=$SERVER_FAILOVER_PERIOD !"
239 log "Starting failover on $SERVERFACET"
241 facet_failover "$SERVERFACET" || exit 1
243 # Check that our client loads are still running during failover.
244 # No application failures should occur.
246 log "==== Checking the clients loads AFTER failover -- failure NOT OK"
247 if ! check_client_loads $NODES_TO_USE; then
248 log "Client load failed during failover. Exiting"
252 # Increment the number of failovers
256 CURRENT_TS=$(date +%s)
257 ELAPSED=$((CURRENT_TS - START_TS))
259 sleep=$((SERVER_FAILOVER_PERIOD-(CURRENT_TS - it_time_start)))
261 # keep count the number of itterations when
262 # time spend to failover and two client loads check exceeded
263 # the value ( SERVER_FAILOVER_PERIOD - MINSLEEP )
264 if [ $sleep -lt $MINSLEEP ]; then
265 reqfail=$((reqfail +1))
266 log "WARNING: failover and two check_client_loads time exceeded SERVER_FAILOVER_PERIOD - MINSLEEP !
267 Failed to load the filesystem with I/O for a minimum period of $MINSLEEP $reqfail times ( REQFAIL=$REQFAIL ).
268 This iteration, the load was only applied for sleep=$sleep seconds.
269 Estimated max recovery time : $max_recov_time
270 Probably the hardware is taking excessively long to boot.
271 Try to increase SERVER_FAILOVER_PERIOD (current is $SERVER_FAILOVER_PERIOD), bug 20918"
272 [ $reqfail -gt $REQFAIL ] && exit 6
275 log "$SERVERFACET has failed over ${!var} times, and counting..."
277 if [ $((ELAPSED + sleep)) -gt $DURATION ]; then
281 if [ $sleep -gt 0 ]; then
282 echo "sleeping $sleep seconds ... "