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 { FAIL_ON_ERROR=true skip_env "$0 Empty SHARED_DIRECTORY" && exit 0; }
32 [ -n "$CLIENTS" ] || \
33 { FAIL_ON_ERROR=true skip_env "$0 Need two or more remote clients" && exit 0; }
35 [ $CLIENTCOUNT -ge 3 ] || \
36 { FAIL_ON_ERROR=true skip_env "$0 Need two or more remote clients, have $((CLIENTCOUNT - 1))" && exit 0; }
38 END_RUN_FILE=${END_RUN_FILE:-$SHARED_DIRECTORY/end_run_file}
39 LOAD_PID_FILE=${LOAD_PID_FILE:-$TMP/client-load.pid}
41 remote_mds_nodsh && skip "remote MDS with nodsh" && exit 0
42 remote_ost_nodsh && skip "remote OST with nodsh" && exit 0
46 check_and_setup_lustre
47 rm -rf $DIR/[df][0-9]*
49 max_recov_time=$(max_recovery_time)
51 # the test node needs to be insulated from a lustre failure as much as possible,
52 # so not even loading the lustre modules is ideal.
54 # -- remove hostname from clients list
55 zconf_umount $(hostname) $MOUNT
56 NODES_TO_USE=${NODES_TO_USE:-$CLIENTS}
57 NODES_TO_USE=$(exclude_items_from_list $NODES_TO_USE $(hostname))
59 check_progs_installed $NODES_TO_USE ${CLIENT_LOADS[@]}
61 MDTS=$(get_facets MDS)
62 OSTS=$(get_facets OST)
64 ERRORS_OK="" # No application failures should occur during this test.
65 FLAVOR=${FLAVOR:-"MDS"}
67 if [ "$FLAVOR" == "MDS" ]; then
73 if [ "$SLOW" = "no" ]; then
74 DURATION=${DURATION:-$((60 * 30))}
75 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 5))}
77 DURATION=${DURATION:-$((60 * 60 * 24))}
78 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 10))} # 10 minutes
83 vmstatLOG=${TESTSUITELOG}_$(basename $0 .sh).vmstat
85 server_numfailovers () {
87 local var=${facet}_numfailovers
90 [[ ${!var} ]] && val=${!var}
94 servers_numfailovers () {
98 for facet in $MDTS ${OSTS//,/ }; do
99 echo "$facet: $(server_numfailovers $facet) times"
103 summary_and_cleanup () {
109 # Having not empty END_RUN_FILE means the failed loads only
110 if [ -s $END_RUN_FILE ]; then
111 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
114 read END_RUN_NODE < $END_RUN_FILE
116 # a client load will end (i.e. fail) if it finds
117 # the end run file. that does not mean that that client load
118 # actually failed though. the first node in the END_RUN_NODE is
119 # the one we are really interested in.
120 if [ -n "$END_RUN_NODE" ]; then
121 var=$(node_var_name $END_RUN_NODE)_load
122 echo "Client load failed on node $END_RUN_NODE"
124 echo "client $END_RUN_NODE load stdout and debug files :
125 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}
126 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}.debug"
131 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
132 echo "$0" >> $END_RUN_FILE
134 [ $rc -eq 0 ] || result=FAIL
136 log "Duration: $DURATION
137 Server failover period: $SERVER_FAILOVER_PERIOD seconds
138 Exited after: $ELAPSED seconds
139 Number of failovers before exit:
140 $(servers_numfailovers)
141 Status: $result: rc=$rc"
143 # stop the vmstats on the OSTs
144 if [ "$VMSTAT" ]; then
145 do_nodes $(comma_list $(osts_nodes)) "test -f /tmp/vmstat.pid && \
146 { kill -s TERM \$(cat /tmp/vmstat.pid); rm -f /tmp/vmstat.pid; \
147 gzip -f9 $vmstatLOG-\$(hostname); }"
150 # make sure the client loads die
151 do_nodes $NODES_TO_USE "set -x; test -f $LOAD_PID_FILE && \
152 { kill -s TERM \$(cat $LOAD_PID_FILE) || true; }"
154 # and free up the pdshes that started them, if any are still around
155 if [ -n "$CLIENT_LOAD_PIDS" ]; then
156 kill $CLIENT_LOAD_PIDS || true
158 kill -9 $CLIENT_LOAD_PIDS || true
160 if [ $rc -ne 0 ]; then
161 # we are interested in only on failed clients and servers
162 local failedclients=$(cat $END_RUN_FILE | grep -v $0)
163 # FIXME: need ostfailover-s nodes also for FLAVOR=OST
164 local product=$(gather_logs $(comma_list $(osts_nodes) \
165 $mds_HOST $mdsfailover_HOST $failedclients))
166 echo logs files $product
169 [ $rc -eq 0 ] && zconf_mount $(hostname) $MOUNT
177 log "-----============= $0 starting =============-----"
179 trap summary_and_cleanup EXIT INT
184 if [ "$VMSTAT" ]; then
185 do_nodes $(comma_list $(osts_nodes)) "vmstat 1 > $vmstatLOG-\$(hostname) 2>/dev/null </dev/null & echo \$! > /tmp/vmstat.pid"
188 # Start client loads.
189 start_client_loads $NODES_TO_USE
191 echo clients load pids:
192 if ! do_nodesv $NODES_TO_USE "cat $LOAD_PID_FILE"; then
196 MINSLEEP=${MINSLEEP:-120}
197 REQFAIL_PERCENT=${REQFAIL_PERCENT:-3} # bug17839 comment 62
198 REQFAIL=${REQFAIL:-$(( DURATION / SERVER_FAILOVER_PERIOD * REQFAIL_PERCENT / 100))}
205 while [ $ELAPSED -lt $DURATION -a ! -e $END_RUN_FILE ]; do
207 # In order to perform the
208 # expected number of failovers, we need to account the following :
209 # 1) the time that has elapsed during the client load checking
210 # 2) time takes for failover
212 it_time_start=$(date +%s)
214 SERVERFACET=$(get_random_entry $SERVERS)
215 var=${SERVERFACET}_numfailovers
217 # Check that our client loads are still running. If any have died,
218 # that means they have died outside of recovery, which is unacceptable.
220 log "==== Checking the clients loads BEFORE failover -- failure NOT OK \
221 ELAPSED=$ELAPSED DURATION=$DURATION PERIOD=$SERVER_FAILOVER_PERIOD"
223 if ! check_client_loads $NODES_TO_USE; then
227 log "Wait $SERVERFACET recovery complete before doing next failover ...."
229 if ! wait_recovery_complete $SERVERFACET ; then
230 echo "$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)) -ge $DURATION ]; then
281 if [ $sleep -gt 0 ]; then
282 echo "sleeping $sleep seconds ... "