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/$NAME.sh}
18 TESTSUITELOG=${TESTSUITELOG:-$TMP/$(basename $0 .sh)}
19 DEBUGLOG=$TESTSUITELOG.debug
24 echo "--- env ---" >&2
26 echo "--- env ---" >&2
29 [ "$SHARED_DIRECTORY" ] || \
30 { skip "$0: Empty SHARED_DIRECTORY" && exit 0; }
32 [ -n "$CLIENTS" ] || { skip "$0 Need two or more remote clients" && exit 0; }
33 [ $CLIENTCOUNT -ge 3 ] || \
34 { skip "$0 Need two or more remote clients, have $CLIENTCOUNT" && exit 0; }
36 END_RUN_FILE=${END_RUN_FILE:-$SHARED_DIRECTORY/end_run_file}
37 LOAD_PID_FILE=${LOAD_PID_FILE:-$TMP/client-load.pid}
39 remote_mds_nodsh && skip "remote MDS with nodsh" && exit 0
40 remote_ost_nodsh && skip "remote OST with nodsh" && exit 0
44 check_and_setup_lustre
45 rm -rf $DIR/[df][0-9]*
47 max_recov_time=$(max_recovery_time)
49 # the test node needs to be insulated from a lustre failure as much as possible,
50 # so not even loading the lustre modules is ideal.
52 # -- remove hostname from clients list
53 zconf_umount $(hostname) $MOUNT
54 NODES_TO_USE=${NODES_TO_USE:-$CLIENTS}
55 NODES_TO_USE=$(exclude_items_from_list $NODES_TO_USE $(hostname))
57 check_progs_installed $NODES_TO_USE ${CLIENT_LOADS[@]}
59 MDTS=$(get_facets MDS)
60 OSTS=$(get_facets OST)
62 ERRORS_OK="" # No application failures should occur during this test.
63 FLAVOR=${FLAVOR:-"MDS"}
65 if [ "$FLAVOR" == "MDS" ]; then
71 if [ "$SLOW" = "no" ]; then
72 DURATION=${DURATION:-$((60 * 30))}
73 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 5))}
75 DURATION=${DURATION:-$((60 * 60 * 24))}
76 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 10))} # 10 minutes
81 vmstatLOG=${TESTSUITELOG}_$(basename $0 .sh).vmstat
83 server_numfailovers () {
85 local var=${facet}_numfailovers
88 [[ ${!var} ]] && val=${!var}
92 servers_numfailovers () {
96 for facet in $MDTS ${OSTS//,/ }; do
97 echo "$facet: $(server_numfailovers $facet) times"
101 summary_and_cleanup () {
107 # Having not empty END_RUN_FILE means the failed loads only
108 if [ -s $END_RUN_FILE ]; then
109 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
112 read END_RUN_NODE < $END_RUN_FILE
114 # a client load will end (i.e. fail) if it finds
115 # the end run file. that does not mean that that client load
116 # actually failed though. the first node in the END_RUN_NODE is
117 # the one we are really interested in.
118 if [ -n "$END_RUN_NODE" ]; then
119 var=$(client_var_name $END_RUN_NODE)_load
120 echo "Client load failed on node $END_RUN_NODE"
122 echo "client $END_RUN_NODE load stdout and debug files :
123 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}
124 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}.debug"
129 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
130 echo "$0" >> $END_RUN_FILE
132 [ $rc -eq 0 ] || result=FAIL
134 log "Duraion: $DURATION
135 Server failover period: $SERVER_FAILOVER_PERIOD seconds
136 Exited after: $ELAPSED seconds
137 Number of failovers before exit:
138 $(servers_numfailovers)
139 Status: $result: rc=$rc"
141 # stop the vmstats on the OSTs
142 if [ "$VMSTAT" ]; then
143 do_nodes $(comma_list $(osts_nodes)) "test -f /tmp/vmstat.pid && \
144 { kill -s TERM \$(cat /tmp/vmstat.pid); rm -f /tmp/vmstat.pid; \
145 gzip -f9 $vmstatLOG-\$(hostname); }"
148 # make sure the client loads die
149 do_nodes $NODES_TO_USE "set -x; test -f $LOAD_PID_FILE && \
150 { kill -s TERM \$(cat $LOAD_PID_FILE) || true; }"
152 # and free up the pdshes that started them, if any are still around
153 if [ -n "$CLIENT_LOAD_PIDS" ]; then
154 kill $CLIENT_LOAD_PIDS || true
156 kill -9 $CLIENT_LOAD_PIDS || true
158 if [ $rc -ne 0 ]; then
159 # we are interested in only on failed clients and servers
160 local failedclients=$(cat $END_RUN_FILE | grep -v $0)
161 # FIXME: need ostfailover-s nodes also for FLAVOR=OST
162 local product=$(gather_logs $(comma_list $(osts_nodes) \
163 $mds_HOST $mdsfailover_HOST $failedclients))
164 echo logs files $product
167 [ $rc -eq 0 ] && zconf_mount $(hostname) $MOUNT
175 log "-----============= $0 starting =============-----"
177 trap summary_and_cleanup EXIT INT
182 if [ "$VMSTAT" ]; then
183 do_nodes $(comma_list $(osts_nodes)) "vmstat 1 > $vmstatLOG-\$(hostname) 2>/dev/null </dev/null & echo \$! > /tmp/vmstat.pid"
186 # Start client loads.
187 start_client_loads $NODES_TO_USE
189 echo clients load pids:
190 if ! do_nodes $NODES_TO_USE "set -x; echo \$(hostname): && cat $LOAD_PID_FILE"; then
191 if [ -e $DEBUGLOG ]; then
198 MINSLEEP=${MINSLEEP:-120}
199 REQFAIL_PERCENT=${REQFAIL_PERCENT:-3} # bug17839 comment 62
200 REQFAIL=${REQFAIL:-$(( DURATION / SERVER_FAILOVER_PERIOD * REQFAIL_PERCENT / 100))}
207 while [ $ELAPSED -lt $DURATION -a ! -e $END_RUN_FILE ]; do
209 # In order to perform the
210 # expected number of failovers, we need to account the following :
211 # 1) the time that has elapsed during the client load checking
212 # 2) time takes for failover
214 it_time_start=$(date +%s)
216 SERVERFACET=$(get_random_entry $SERVERS)
217 var=${SERVERFACET}_numfailovers
219 # Check that our client loads are still running. If any have died,
220 # that means they have died outside of recovery, which is unacceptable.
222 log "==== Checking the clients loads BEFORE failover -- failure NOT OK \
223 ELAPSED=$ELAPSED DURATION=$DURATION PERIOD=$SERVER_FAILOVER_PERIOD"
225 if ! check_client_loads $NODES_TO_USE; then
229 log "Wait $SERVERFACET recovery complete before doing next failover ...."
230 if [[ $(server_numfailovers $SERVERFACET) != 0 ]]; then
231 if ! wait_recovery_complete $SERVERFACET ; then
232 echo "$SERVERFACET recovery is not completed!"
237 log "Checking clients are in FULL state before doing next failover"
238 if ! wait_clients_import_state $NODES_TO_USE $SERVERFACET FULL; then
239 echo "Clients import not FULL, please consider to increase SERVER_FAILOVER_PERIOD=$SERVER_FAILOVER_PERIOD !"
242 log "Starting failover on $SERVERFACET"
244 facet_failover "$SERVERFACET" || exit 1
246 # Check that our client loads are still running during failover.
247 # No application failures should occur.
249 log "==== Checking the clients loads AFTER failover -- failure NOT OK"
250 if ! check_client_loads $NODES_TO_USE; then
251 log "Client load failed during failover. Exiting"
255 # Increment the number of failovers
259 CURRENT_TS=$(date +%s)
260 ELAPSED=$((CURRENT_TS - START_TS))
262 sleep=$((SERVER_FAILOVER_PERIOD-(CURRENT_TS - it_time_start)))
264 # keep count the number of itterations when
265 # time spend to failover and two client loads check exceeded
266 # the value ( SERVER_FAILOVER_PERIOD - MINSLEEP )
267 if [ $sleep -lt $MINSLEEP ]; then
268 reqfail=$((reqfail +1))
269 log "WARNING: failover and two check_client_loads time exceeded SERVER_FAILOVER_PERIOD - MINSLEEP !
270 Failed to load the filesystem with I/O for a minimum period of $MINSLEEP $reqfail times ( REQFAIL=$REQFAIL ).
271 This iteration, the load was only applied for sleep=$sleep seconds.
272 Estimated max recovery time : $max_recov_time
273 Probably the hardware is taking excessively long to boot.
274 Try to increase SERVER_FAILOVER_PERIOD (current is $SERVER_FAILOVER_PERIOD), bug 20918"
275 [ $reqfail -gt $REQFAIL ] && exit 6
278 log "$SERVERFACET has failed over ${!var} times, and counting..."
280 if [ $((ELAPSED + sleep)) -gt $DURATION ]; then
284 if [ $sleep -gt 0 ]; then
285 echo "sleeping $sleep seconds ... "