3 # client failure does not affect other clients
5 # Start load on clients (each client works on it's own directory).
6 # At defined (5-10 minutes) interval fail one random client and then fail mds.
7 # Reintegrate failed client after recovery completed,
8 # application errors are allowed for that client but not on other clients.
9 # 10 minute intervals and verify that no application errors occur.
11 # Test runs one of CLIENT_LOAD progs on remote clients.
13 LUSTRE=${LUSTRE:-`dirname $0`/..}
15 CLEANUP=${CLEANUP:-""}
16 . $LUSTRE/tests/test-framework.sh
20 . ${CONFIG:=$LUSTRE/tests/cfg/$NAME.sh}
22 TESTSUITELOG=${TESTSUITELOG:-$TMP/$(basename $0 .sh)}
23 DEBUGLOG=$TESTSUITELOG.debug
28 echo "--- env ---" >&2
30 echo "--- env ---" >&2
33 [ "$SHARED_DIRECTORY" ] || \
34 { skip "$0: Empty SHARED_DIRECTORY" && exit 0; }
36 [ -n "$CLIENTS" ] || { skip "$0 Need two or more remote clients" && exit 0; }
37 [ $CLIENTCOUNT -ge 3 ] || \
38 { skip "$0 Need two or more clients, have $CLIENTCOUNT" && exit 0; }
40 END_RUN_FILE=${END_RUN_FILE:-$SHARED_DIRECTORY/end_run_file}
41 LOAD_PID_FILE=${LOAD_PID_FILE:-$TMP/client-load.pid}
43 remote_mds_nodsh && skip "remote MDS with nodsh" && exit 0
47 check_and_setup_lustre
48 rm -rf $DIR/[df][0-9]*
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)
62 if [ "$SLOW" = "no" ]; then
63 DURATION=${DURATION:-$((60 * 30))}
64 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 5))}
66 DURATION=${DURATION:-$((60 * 60 * 24))}
67 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 10))} # 10 minutes
72 vmstatLOG=${TESTSUITELOG}_$(basename $0 .sh).vmstat
78 for facet in $MDTS ${failed_clients//,/ }; do
82 echo "$facet failed over $val times"
87 # list is comma separated
91 do_nodes $list "node=\\\$(hostname)
93 log=${TESTSUITELOG}_run_${!var}.sh-\\\$node.debug
94 if [ -e \\\$log ] ; then
95 echo Node \\\$node debug log:
100 summary_and_cleanup () {
106 # Having not empty END_RUN_FILE means the failed loads only
107 if [ -s $END_RUN_FILE ]; then
108 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
111 read END_RUN_NODE < $END_RUN_FILE
113 # a client load will end (i.e. fail) if it finds
114 # the end run file. that does not mean that that client load
115 # actually failed though. the first node in the END_RUN_NODE is
116 # the one we are really interested in.
117 if [ -n "$END_RUN_NODE" ]; then
118 var=$(client_var_name $END_RUN_NODE)_load
119 echo "Client load failed on node $END_RUN_NODE"
121 echo "client $END_RUN_NODE load stdout and debug files :
122 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}
123 ${TESTSUITELOG}_run_${!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:
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
159 if [ $rc -ne 0 ]; then
160 print_logs $NODES_TO_USE
161 # we are interested in only on failed clients and servers
162 local failedclients=$(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 $(mdts_nodes) $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_nodes $NODES_TO_USE "set -x; echo \$(hostname): && cat $LOAD_PID_FILE"; then
193 if [ -e $DEBUGLOG ]; then
203 MINSLEEP=${MINSLEEP:-120}
204 REQFAIL_PERCENT=${REQFAIL_PERCENT:-3} # bug17839 comment 62
205 REQFAIL=${REQFAIL:-$(( DURATION / SERVER_FAILOVER_PERIOD * REQFAIL_PERCENT / 100))}
209 # This is used for FAIL_CLIENT only
211 while [ $ELAPSED -lt $DURATION -a ! -e $END_RUN_FILE ]; do
213 # In order to perform the
214 # expected number of failovers, we need to account the following :
215 # 1) the time that has elapsed during the client load checking
216 # 2) time takes for failover
218 it_time_start=$(date +%s)
220 FAIL_CLIENT=$(get_random_entry $NODES_TO_USE)
221 client_var=$(client_var_name $FAIL_CLIENT)_nums
223 # store the list of failed clients
224 # lists are comma separated
225 failed_clients=$(expand_list $failed_clients $FAIL_CLIENT)
227 SERVERFACET=$(get_random_entry $MDTS)
228 var=${SERVERFACET}_nums
230 # Check that our client loads are still running. If any have died,
231 # that means they have died outside of recovery, which is unacceptable.
233 log "==== Checking the clients loads BEFORE failover -- failure NOT OK \
234 ELAPSED=$ELAPSED DURATION=$DURATION PERIOD=$SERVER_FAILOVER_PERIOD"
236 if ! check_client_loads $NODES_TO_USE; then
240 log "FAIL CLIENT $FAIL_CLIENT ... "
241 shutdown_client $FAIL_CLIENT
243 log "Starting failover on $SERVERFACET"
245 facet_failover "$SERVERFACET" || exit 1
246 if ! wait_recovery_complete $SERVERFACET $((TIMEOUT * 10)); then
247 echo "$SERVERFACET recovery is not completed!"
251 boot_node $FAIL_CLIENT
252 echo "Reintegrating $FAIL_CLIENT"
253 zconf_mount $FAIL_CLIENT $MOUNT || exit $?
255 # Increment the number of failovers
258 val=$((${!client_var} + 1))
259 eval $client_var=$val
261 # load script on failed clients could create END_RUN_FILE
262 # We shuold remove it and ignore the failure if this
263 # file contains the failed client only.
264 # We can not use ERRORS_OK when start all loads at the start of this script
265 # because the application errors allowed for random failed client only, but
266 # not for all clients.
267 if [ -e $END_RUN_FILE ]; then
268 read END_RUN_NODE < $END_RUN_FILE
269 [[ $END_RUN_NODE = $FAIL_CLIENT ]] &&
270 rm -f $END_RUN_FILE || exit 13
273 restart_client_loads $FAIL_CLIENT $ERRORS_OK || exit $?
275 # Check that not failed clients loads are still running.
276 # No application failures should occur on clients that was not failed.
278 log "==== Checking the clients loads AFTER failed client reintegrated -- failure NOT OK"
279 if ! ERRORS_OK= check_client_loads $(exclude_items_from_list $NODES_TO_USE $FAIL_CLIENT); then
280 log "Client load failed. Exiting"
284 CURRENT_TS=$(date +%s)
285 ELAPSED=$((CURRENT_TS - START_TS))
287 sleep=$((SERVER_FAILOVER_PERIOD-(CURRENT_TS - it_time_start)))
289 # keep count the number of itterations when
290 # time spend to failover and two client loads check exceeded
291 # the value ( SERVER_FAILOVER_PERIOD - MINSLEEP )
292 if [ $sleep -lt $MINSLEEP ]; then
293 reqfail=$((reqfail +1))
294 log "WARNING: failover, client reintegration and check_client_loads time
295 exceeded SERVER_FAILOVER_PERIOD - MINSLEEP !
296 Failed to meet interval $reqfail times ( REQFAIL=$REQFAIL ); have sleep=$sleep"
297 [ $reqfail -gt $REQFAIL ] && exit 6
300 log " Number of failovers:
301 $(numfailovers) and counting..."
303 if [ $((ELAPSED + sleep)) -gt $DURATION ]; then
307 if [ $sleep -gt 0 ]; then
308 echo "sleeping $sleep seconds ... "