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.
16 LUSTRE=${LUSTRE:-$(dirname $0)/..}
17 . $LUSTRE/tests/test-framework.sh
21 # bug number for skipped test:
22 ALWAYS_EXCEPT="$RECOVERY_RANDOM_SCALE_EXCEPT"
23 # UPDATE THE COMMENT ABOVE WITH BUG NUMBERS WHEN CHANGING ALWAYS_EXCEPT!
27 remote_mds_nodsh && skip_env "remote MDS with nodsh" && exit 0
28 remote_ost_nodsh && skip_env "remote OST with nodsh" && exit 0
30 [ $CLIENTCOUNT -lt 3 ] &&
31 skip_env "need three or more clients" && exit 0
33 if [ -z "$SHARED_DIRECTORY" ] || ! check_shared_dir $SHARED_DIRECTORY; then
34 skip_env "SHARED_DIRECTORY should be specified with a shared directory \
35 which is accessable on all of the nodes"
39 [[ $FAILURE_MODE = SOFT ]] && \
40 log "WARNING: $0 is not functional with FAILURE_MODE = SOFT, bz22797"
42 # Application failures are allowed for the failed client
43 # but not for other clients.
46 if [ "$SLOW" = "no" ]; then
47 DURATION=${DURATION:-$((60 * 30))}
49 DURATION=${DURATION:-$((60 * 60 * 24))}
51 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 10))} # 10 minutes
53 MINSLEEP=${MINSLEEP:-120}
54 REQFAIL_PERCENT=${REQFAIL_PERCENT:-3} # bug17839 comment 62
55 # round up the result of integer division: C=(A + (B - 1)) / B
56 REQFAIL=${REQFAIL:-$(((DURATION * REQFAIL_PERCENT + (SERVER_FAILOVER_PERIOD *
57 100 - 1 )) / SERVER_FAILOVER_PERIOD / 100))}
59 END_RUN_FILE=${END_RUN_FILE:-$SHARED_DIRECTORY/end_run_file}
60 LOAD_PID_FILE=${LOAD_PID_FILE:-$TMP/client-load.pid}
61 VMSTAT_PID_FILE=${VMSTAT_PID_FILE:-$TMP/vmstat.pid}
67 for facet in ${MDTS//,/ } ${FAILED_CLIENTS//,/ }; do
68 var=$(node_var_name $facet)_nums
71 echo "$facet failed over $val times"
76 summary_and_cleanup () {
80 # Having not empty END_RUN_FILE means the failed loads only
81 if [ -s $END_RUN_FILE ]; then
82 print_end_run_file $END_RUN_FILE
86 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
87 echo "$0" >> $END_RUN_FILE
89 [ $rc -eq 0 ] || result=FAIL
91 log "Duration: $DURATION
92 Server failover period: $SERVER_FAILOVER_PERIOD seconds
93 Exited after: $ELAPSED seconds
94 Number of failovers before exit:
96 Status: $result: rc=$rc"
98 # stop vmstat on OSS nodes
99 [ "$VMSTAT" ] && stop_process $(comma_list $(osts_nodes)) $VMSTAT_PID_FILE
101 # stop the client loads
102 stop_client_loads $NODES_TO_USE $LOAD_PID_FILE
104 if [ $rc -ne 0 ]; then
105 # we are interested in only on failed clients and servers
106 local failedclients=$(cat $END_RUN_FILE | grep -v $0)
107 gather_logs $(comma_list $(all_server_nodes) $failedclients)
113 ################################## Main Flow ###################################
114 check_and_setup_lustre
115 rm -rf $DIR/[Rdfs][0-9]*
117 MAX_RECOV_TIME=$(max_recovery_time)
119 # The test node needs to be insulated from a lustre failure as much as possible,
120 # so not even loading the lustre modules is ideal.
122 # -- remove hostname from clients list
123 zconf_umount $HOSTNAME $MOUNT
124 NODES_TO_USE=${NODES_TO_USE:-$CLIENTS}
125 NODES_TO_USE=$(exclude_items_from_list $NODES_TO_USE $HOSTNAME)
127 check_progs_installed $NODES_TO_USE ${CLIENT_LOADS[@]}
129 MDTS=$(get_facets MDS)
131 # Fail a random client and then failover a random MDS.
132 test_fail_client_mds() {
138 trap summary_and_cleanup EXIT INT
140 # start vmstat on OSS nodes
141 [ "$VMSTAT" ] && start_vmstat $(comma_list $(osts_nodes)) $VMSTAT_PID_FILE
145 start_client_loads $NODES_TO_USE
147 echo client loads pids:
148 do_nodesv $NODES_TO_USE "cat $LOAD_PID_FILE" || exit 3
154 local start_ts=$(date +%s)
155 local current_ts=$start_ts
157 while [ $ELAPSED -lt $DURATION -a ! -e $END_RUN_FILE ]; do
158 # In order to perform the
159 # expected number of failovers, we need to account the following:
160 # 1) the time that has elapsed during the client load checking
161 # 2) time takes for failover
162 it_time_start=$(date +%s)
164 fail_client=$(get_random_entry $NODES_TO_USE)
165 client_var=$(node_var_name $fail_client)_nums
167 # store the list of failed clients
168 # lists are comma separated
169 FAILED_CLIENTS=$(expand_list $FAILED_CLIENTS $fail_client)
171 serverfacet=$(get_random_entry $MDTS)
172 var=$(node_var_name $serverfacet)_nums
174 # Check that our client loads are still running. If any have died,
175 # that means they have died outside of recovery, which is unacceptable.
176 log "==== Checking the clients loads BEFORE failover -- failure NOT OK \
177 ELAPSED=$ELAPSED DURATION=$DURATION PERIOD=$SERVER_FAILOVER_PERIOD"
178 check_client_loads $NODES_TO_USE || exit 4
180 log "FAIL CLIENT $fail_client..."
181 shutdown_client $fail_client
183 log "Starting failover on $serverfacet"
184 facet_failover "$serverfacet" || exit 1
186 if ! wait_recovery_complete $serverfacet; then
187 echo "$serverfacet recovery is not completed!"
191 boot_node $fail_client
192 echo "Reintegrating $fail_client"
193 zconf_mount $fail_client $MOUNT || exit $?
194 client_up $fail_client || exit $?
196 # Increment the number of failovers
199 val=$((${!client_var} + 1))
200 eval $client_var=$val
202 # load script on failed clients could create END_RUN_FILE
203 # We shuold remove it and ignore the failure if this
204 # file contains the failed client only.
205 # We can not use ERRORS_OK when start all loads at the start of
206 # this script because the application errors allowed for random
207 # failed client only, but not for all clients.
208 if [ -e $END_RUN_FILE ]; then
210 read end_run_node < $END_RUN_FILE
211 [[ $end_run_node = $fail_client ]] &&
212 rm -f $END_RUN_FILE || exit 13
215 restart_client_loads $fail_client $ERRORS_OK || exit $?
217 # Check that not failed clients loads are still running.
218 # No application failures should occur on clients that were not failed.
219 log "==== Checking the clients loads AFTER failed client reintegrated \
221 if ! ERRORS_OK= check_client_loads \
222 $(exclude_items_from_list $NODES_TO_USE $fail_client); then
223 log "Client load failed. Exiting..."
227 current_ts=$(date +%s)
228 ELAPSED=$((current_ts - start_ts))
229 sleep=$((SERVER_FAILOVER_PERIOD - (current_ts - it_time_start)))
231 # Keep counting the number of iterations when
232 # time spent to failover and two client loads check exceeded
233 # the value ( SERVER_FAILOVER_PERIOD - MINSLEEP ).
234 if [ $sleep -lt $MINSLEEP ]; then
235 reqfail=$((reqfail + 1))
236 log "WARNING: failover, client reintegration and \
237 check_client_loads time exceeded SERVER_FAILOVER_PERIOD - MINSLEEP!
238 Failed to load the filesystem with I/O for a minimum period of \
239 $MINSLEEP $reqfail times ( REQFAIL=$REQFAIL ).
240 This iteration, the load was only applied for sleep=$sleep seconds.
241 Estimated max recovery time : $MAX_RECOV_TIME
242 Probably the hardware is taking excessively long time to boot.
243 Try to increase SERVER_FAILOVER_PERIOD (current is $SERVER_FAILOVER_PERIOD), \
245 [ $reqfail -gt $REQFAIL ] && exit 6
248 log "Number of failovers:
249 $(numfailovers) and counting..."
251 [ $((ELAPSED + sleep)) -ge $DURATION ] && break
253 if [ $sleep -gt 0 ]; then
254 echo "sleeping $sleep seconds... "
260 run_test fail_client_mds "fail client, then failover MDS"
262 zconf_mount $HOSTNAME $MOUNT || error "mount $MOUNT on $HOSTNAME failed"
263 client_up || error "start client on $HOSTNAME failed"
266 check_and_cleanup_lustre