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 { FAIL_ON_ERROR=true skip_env "$0 Empty SHARED_DIRECTORY" && exit 0; }
33 [ -n "$CLIENTS" ] || \
34 { FAIL_ON_ERROR=true skip_env "$0 Need two or more remote clients" && exit 0; }
36 [ $CLIENTCOUNT -ge 3 ] || \
37 { FAIL_ON_ERROR=true skip_env "$0 Need two or more remote clients, have $((CLIENTCOUNT - 1))" && exit 0; }
39 END_RUN_FILE=${END_RUN_FILE:-$SHARED_DIRECTORY/end_run_file}
40 LOAD_PID_FILE=${LOAD_PID_FILE:-$TMP/client-load.pid}
42 remote_mds_nodsh && skip "remote MDS with nodsh" && exit 0
43 remote_ost_nodsh && skip "remote OST with nodsh" && exit 0
47 check_and_setup_lustre
48 rm -rf $DIR/[df][0-9]*
50 max_recov_time=$(max_recovery_time)
52 # the test node needs to be insulated from a lustre failure as much as possible,
53 # so not even loading the lustre modules is ideal.
55 # -- remove hostname from clients list
56 zconf_umount $(hostname) $MOUNT
57 NODES_TO_USE=${NODES_TO_USE:-$CLIENTS}
58 NODES_TO_USE=$(exclude_items_from_list $NODES_TO_USE $(hostname))
60 check_progs_installed $NODES_TO_USE ${CLIENT_LOADS[@]}
62 MDTS=$(get_facets MDS)
63 OSTS=$(get_facets OST)
65 ERRORS_OK="" # No application failures should occur during this test.
66 FLAVOR=${FLAVOR:-"MDS"}
68 if [ "$FLAVOR" == "MDS" ]; then
74 if [ "$SLOW" = "no" ]; then
75 DURATION=${DURATION:-$((60 * 30))}
76 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 5))}
78 DURATION=${DURATION:-$((60 * 60 * 24))}
79 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 10))} # 10 minutes
84 vmstatLOG=${TESTSUITELOG}_$(basename $0 .sh).vmstat
86 server_numfailovers () {
88 local var=${facet}_numfailovers
91 [[ ${!var} ]] && val=${!var}
95 servers_numfailovers () {
99 for facet in ${MDTS//,/ } ${OSTS//,/ }; do
100 echo "$facet: $(server_numfailovers $facet) times"
104 summary_and_cleanup () {
110 # Having not empty END_RUN_FILE means the failed loads only
111 if [ -s $END_RUN_FILE ]; then
112 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
115 read END_RUN_NODE < $END_RUN_FILE
117 # a client load will end (i.e. fail) if it finds
118 # the end run file. that does not mean that that client load
119 # actually failed though. the first node in the END_RUN_NODE is
120 # the one we are really interested in.
121 if [ -n "$END_RUN_NODE" ]; then
122 var=$(client_var_name $END_RUN_NODE)_load
123 echo "Client load failed on node $END_RUN_NODE"
125 echo "client $END_RUN_NODE load stdout and debug files :
126 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}
127 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}.debug"
132 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
133 echo "$0" >> $END_RUN_FILE
135 [ $rc -eq 0 ] || result=FAIL
137 log "Duration: $DURATION
138 Server failover period: $SERVER_FAILOVER_PERIOD seconds
139 Exited after: $ELAPSED seconds
140 Number of failovers before exit:
141 $(servers_numfailovers)
142 Status: $result: rc=$rc"
144 # stop the vmstats on the OSTs
145 if [ "$VMSTAT" ]; then
146 do_nodes $(comma_list $(osts_nodes)) "test -f /tmp/vmstat.pid && \
147 { kill -s TERM \$(cat /tmp/vmstat.pid); rm -f /tmp/vmstat.pid; \
148 gzip -f9 $vmstatLOG-\$(hostname); }"
151 # make sure the client loads die
152 do_nodes $NODES_TO_USE "set -x; test -f $LOAD_PID_FILE && \
153 { kill -s TERM \$(cat $LOAD_PID_FILE) || true; }"
155 # and free up the pdshes that started them, if any are still around
156 if [ -n "$CLIENT_LOAD_PIDS" ]; then
157 kill $CLIENT_LOAD_PIDS || true
159 kill -9 $CLIENT_LOAD_PIDS || true
161 if [ $rc -ne 0 ]; then
162 # we are interested in only on failed clients and servers
163 local failedclients=$(cat $END_RUN_FILE | grep -v $0)
164 # FIXME: need ostfailover-s nodes also for FLAVOR=OST
165 local product=$(gather_logs $(comma_list $(osts_nodes) \
166 $(mdts_nodes) $mdsfailover_HOST $failedclients))
167 echo logs files $product
170 [ $rc -eq 0 ] && zconf_mount $(hostname) $MOUNT
178 log "-----============= $0 starting =============-----"
180 trap summary_and_cleanup EXIT INT
185 if [ "$VMSTAT" ]; then
186 do_nodes $(comma_list $(osts_nodes)) "vmstat 1 > $vmstatLOG-\$(hostname) 2>/dev/null </dev/null & echo \$! > /tmp/vmstat.pid"
189 # Start client loads.
190 start_client_loads $NODES_TO_USE
192 echo clients load pids:
193 if ! do_nodesv $NODES_TO_USE "cat $LOAD_PID_FILE"; then
197 MINSLEEP=${MINSLEEP:-120}
198 REQFAIL_PERCENT=${REQFAIL_PERCENT:-3} # bug17839 comment 62
199 REQFAIL=${REQFAIL:-$(( DURATION / SERVER_FAILOVER_PERIOD * REQFAIL_PERCENT / 100))}
206 while [ $ELAPSED -lt $DURATION -a ! -e $END_RUN_FILE ]; do
208 # In order to perform the
209 # expected number of failovers, we need to account the following :
210 # 1) the time that has elapsed during the client load checking
211 # 2) time takes for failover
213 it_time_start=$(date +%s)
215 SERVERFACET=$(get_random_entry $SERVERS)
216 var=${SERVERFACET}_numfailovers
218 # Check that our client loads are still running. If any have died,
219 # that means they have died outside of recovery, which is unacceptable.
221 log "==== Checking the clients loads BEFORE failover -- failure NOT OK \
222 ELAPSED=$ELAPSED DURATION=$DURATION PERIOD=$SERVER_FAILOVER_PERIOD"
224 if ! check_client_loads $NODES_TO_USE; then
228 log "Wait $SERVERFACET recovery complete before doing next failover ...."
230 if ! wait_recovery_complete $SERVERFACET ; then
231 echo "$SERVERFACET recovery is not completed!"
235 log "Checking clients are in FULL state before doing next failover"
236 if ! wait_clients_import_state $NODES_TO_USE $SERVERFACET FULL; then
237 echo "Clients import not FULL, please consider to increase SERVER_FAILOVER_PERIOD=$SERVER_FAILOVER_PERIOD !"
240 log "Starting failover on $SERVERFACET"
242 facet_failover "$SERVERFACET" || exit 1
244 # Check that our client loads are still running during failover.
245 # No application failures should occur.
247 log "==== Checking the clients loads AFTER failover -- failure NOT OK"
248 if ! check_client_loads $NODES_TO_USE; then
249 log "Client load failed during failover. Exiting"
253 # Increment the number of failovers
257 CURRENT_TS=$(date +%s)
258 ELAPSED=$((CURRENT_TS - START_TS))
260 sleep=$((SERVER_FAILOVER_PERIOD-(CURRENT_TS - it_time_start)))
262 # keep count the number of itterations when
263 # time spend to failover and two client loads check exceeded
264 # the value ( SERVER_FAILOVER_PERIOD - MINSLEEP )
265 if [ $sleep -lt $MINSLEEP ]; then
266 reqfail=$((reqfail +1))
267 log "WARNING: failover and two check_client_loads time exceeded SERVER_FAILOVER_PERIOD - MINSLEEP !
268 Failed to load the filesystem with I/O for a minimum period of $MINSLEEP $reqfail times ( REQFAIL=$REQFAIL ).
269 This iteration, the load was only applied for sleep=$sleep seconds.
270 Estimated max recovery time : $max_recov_time
271 Probably the hardware is taking excessively long to boot.
272 Try to increase SERVER_FAILOVER_PERIOD (current is $SERVER_FAILOVER_PERIOD), bug 20918"
273 [ $reqfail -gt $REQFAIL ] && exit 6
276 log "$SERVERFACET has failed over ${!var} times, and counting..."
278 if [ $((ELAPSED + sleep)) -ge $DURATION ]; then
282 if [ $sleep -gt 0 ]; then
283 echo "sleeping $sleep seconds ... "