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:-$TMP/recovery-random-scale}
23 DEBUGLOG=$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 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
44 check_and_setup_lustre
45 rm -rf $DIR/[df][0-9]*
47 # the test node needs to be insulated from a lustre failure as much as possible,
48 # so not even loading the lustre modules is ideal.
50 # -- remove hostname from clients list
51 zconf_umount $(hostname) $MOUNT
52 NODES_TO_USE=${NODES_TO_USE:-$CLIENTS}
53 NODES_TO_USE=$(exclude_items_from_list $NODES_TO_USE $(hostname))
55 check_progs_installed $NODES_TO_USE ${CLIENT_LOADS[@]}
57 MDTS=$(get_facets MDS)
59 if [ "$SLOW" = "no" ]; then
60 DURATION=${DURATION:-$((60 * 30))}
61 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 5))}
63 DURATION=${DURATION:-$((60 * 60 * 24))}
64 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 10))} # 10 minutes
69 vmstatLOG=${TESTSUITELOG}_$(basename $0 .sh).vmstat
75 for facet in $MDTS ${failed_clients//,/ }; do
79 echo "$facet failed over $val times"
84 # list is comma separated
88 do_nodes $list "node=\\\$(hostname)
90 log=${TESTSUITELOG}_run_${!var}.sh-\\\$node.debug
91 if [ -e \\\$log ] ; then
92 echo Node \\\$node debug log:
97 summary_and_cleanup () {
103 # Having not empty END_RUN_FILE means the failed loads only
104 if [ -s $END_RUN_FILE ]; then
105 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
108 read END_RUN_NODE < $END_RUN_FILE
110 # a client load will end (i.e. fail) if it finds
111 # the end run file. that does not mean that that client load
112 # actually failed though. the first node in the END_RUN_NODE is
113 # the one we are really interested in.
114 if [ -n "$END_RUN_NODE" ]; then
115 var=$(client_var_name $END_RUN_NODE)_load
116 echo "Client load failed on node $END_RUN_NODE"
118 echo "client $END_RUN_NODE load stdout and debug files :
119 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}
120 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}.debug"
126 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
127 echo "$0" >> $END_RUN_FILE
129 [ $rc -eq 0 ] || result=FAIL
131 log "Duraion: $DURATION
132 Server failover period: $SERVER_FAILOVER_PERIOD seconds
133 Exited after: $ELAPSED seconds
134 Number of failovers before exit:
136 Status: $result: rc=$rc"
138 # stop the vmstats on the OSTs
139 if [ "$VMSTAT" ]; then
140 do_nodes $(comma_list $(osts_nodes)) "test -f /tmp/vmstat.pid && \
141 { kill -s TERM \$(cat /tmp/vmstat.pid); rm -f /tmp/vmstat.pid; \
142 gzip -f9 $vmstatLOG-\$(hostname); }"
145 # make sure the client loads die
146 do_nodes $NODES_TO_USE "set -x; test -f $LOAD_PID_FILE && \
147 { kill -s TERM \$(cat $LOAD_PID_FILE) || true; }"
149 # and free up the pdshes that started them, if any are still around
150 if [ -n "$CLIENT_LOAD_PIDS" ]; then
151 kill $CLIENT_LOAD_PIDS || true
153 kill -9 $CLIENT_LOAD_PIDS || true
156 if [ $rc -ne 0 ]; then
157 print_logs $NODES_TO_USE
160 if [ $rc -eq 0 ]; then
161 zconf_mount $(hostname) $MOUNT
163 error "exited with rc=$rc"
171 log "-----============= $0 starting =============-----"
173 trap summary_and_cleanup EXIT # INT
178 if [ "$VMSTAT" ]; then
179 do_nodes $(comma_list $(osts_nodes)) "vmstat 1 > $vmstatLOG-\$(hostname) 2>/dev/null </dev/null & echo \$! > /tmp/vmstat.pid"
182 # Start client loads.
183 start_client_loads $NODES_TO_USE
185 echo clients load pids:
186 if ! do_nodes $NODES_TO_USE "set -x; echo \$(hostname): && cat $LOAD_PID_FILE"; then
187 if [ -e $DEBUGLOG ]; then
197 MINSLEEP=${MINSLEEP:-120}
198 REQFAIL_PERCENT=${REQFAIL_PERCENT:-3} # bug17839 comment 62
199 REQFAIL=${REQFAIL:-$(( DURATION / SERVER_FAILOVER_PERIOD * REQFAIL_PERCENT / 100))}
203 # This is used for FAIL_CLIENT only
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 FAIL_CLIENT=$(get_random_entry $NODES_TO_USE)
215 client_var=$(client_var_name $FAIL_CLIENT)_nums
217 # store the list of failed clients
218 # lists are comma separated
219 failed_clients=$(expand_list $failed_clients $FAIL_CLIENT)
221 SERVERFACET=$(get_random_entry $MDTS)
222 var=${SERVERFACET}_nums
224 # Check that our client loads are still running. If any have died,
225 # that means they have died outside of recovery, which is unacceptable.
227 log "==== Checking the clients loads BEFORE failover -- failure NOT OK \
228 ELAPSED=$ELAPSED DURATION=$DURATION PERIOD=$SERVER_FAILOVER_PERIOD"
230 if ! check_client_loads $NODES_TO_USE; then
234 log "FAIL CLIENT $FAIL_CLIENT ... "
235 shutdown_client $FAIL_CLIENT
237 log "Starting failover on $SERVERFACET"
239 facet_failover "$SERVERFACET" || exit 1
240 if ! wait_recovery_complete $SERVERFACET $((TIMEOUT * 10)); then
241 echo "$SERVERFACET recovery is not completed!"
245 boot_node $FAIL_CLIENT
246 echo "Reintegrating $FAIL_CLIENT"
247 zconf_mount $FAIL_CLIENT $MOUNT || exit $?
249 # Increment the number of failovers
252 val=$((${!client_var} + 1))
253 eval $client_var=$val
255 # load script on failed clients could create END_RUN_FILE
256 # We shuold remove it and ignore the failure if this
257 # file contains the failed client only.
258 # We can not use ERRORS_OK when start all loads at the start of this script
259 # because the application errors allowed for random failed client only, but
260 # not for all clients.
261 if [ -e $END_RUN_FILE ]; then
262 read END_RUN_NODE < $END_RUN_FILE
263 [[ $END_RUN_NODE = $FAIL_CLIENT ]] &&
264 rm -f $END_RUN_FILE || exit 13
267 restart_client_loads $FAIL_CLIENT $ERRORS_OK || exit $?
269 # Check that not failed clients loads are still running.
270 # No application failures should occur on clients that was not failed.
272 log "==== Checking the clients loads AFTER failed client reintegrated -- failure NOT OK"
273 if ! ERRORS_OK= check_client_loads $(exclude_items_from_list $NODES_TO_USE $FAIL_CLIENT); then
274 log "Client load failed. Exiting"
278 CURRENT_TS=$(date +%s)
279 ELAPSED=$((CURRENT_TS - START_TS))
281 sleep=$((SERVER_FAILOVER_PERIOD-(CURRENT_TS - it_time_start)))
283 # keep count the number of itterations when
284 # time spend to failover and two client loads check exceeded
285 # the value ( SERVER_FAILOVER_PERIOD - MINSLEEP )
286 if [ $sleep -lt $MINSLEEP ]; then
287 reqfail=$((reqfail +1))
288 log "WARNING: failover, client reintegration and check_client_loads time
289 exceeded SERVER_FAILOVER_PERIOD - MINSLEEP !
290 Failed to meet interval $reqfail times ( REQFAIL=$REQFAIL ); have sleep=$sleep"
291 [ $reqfail -gt $REQFAIL ] && exit 6
294 log " Number of failovers:
295 $(numfailovers) and counting..."
297 if [ $sleep -gt 0 ]; then
298 echo "sleeping $sleep seconds ... "