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/$NAME.sh}
23 TESTSUITELOG=${TESTSUITELOG:-$TMP/$(basename $0 .sh)}
24 DEBUGLOG=$TESTSUITELOG.debug
29 echo "--- env ---" >&2
31 echo "--- env ---" >&2
34 [ "$SHARED_DIRECTORY" ] || \
35 { skip "$0: Empty SHARED_DIRECTORY" && exit 0; }
37 [ -n "$CLIENTS" ] || { skip "$0 Need two or more remote clients" && exit 0; }
38 [ $CLIENTCOUNT -ge 3 ] || \
39 { skip "$0 Need two or more clients, have $CLIENTCOUNT" && exit 0; }
41 END_RUN_FILE=${END_RUN_FILE:-$SHARED_DIRECTORY/end_run_file}
42 LOAD_PID_FILE=${LOAD_PID_FILE:-$TMP/client-load.pid}
44 remote_mds_nodsh && skip "remote MDS with nodsh" && exit 0
48 check_and_setup_lustre
49 rm -rf $DIR/[df][0-9]*
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)
63 if [ "$SLOW" = "no" ]; then
64 DURATION=${DURATION:-$((60 * 30))}
65 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 5))}
67 DURATION=${DURATION:-$((60 * 60 * 24))}
68 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 10))} # 10 minutes
73 vmstatLOG=${TESTSUITELOG}_$(basename $0 .sh).vmstat
79 for facet in $MDTS ${failed_clients//,/ }; do
83 echo "$facet failed over $val times"
88 # list is comma separated
92 do_nodes $list "node=\\\$(hostname)
94 log=${TESTSUITELOG}_run_${!var}.sh-\\\$node.debug
95 if [ -e \\\$log ] ; then
96 echo Node \\\$node debug log:
101 summary_and_cleanup () {
107 # Having not empty END_RUN_FILE means the failed loads only
108 if [ -s $END_RUN_FILE ]; then
109 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
112 read END_RUN_NODE < $END_RUN_FILE
114 # a client load will end (i.e. fail) if it finds
115 # the end run file. that does not mean that that client load
116 # actually failed though. the first node in the END_RUN_NODE is
117 # the one we are really interested in.
118 if [ -n "$END_RUN_NODE" ]; then
119 var=$(client_var_name $END_RUN_NODE)_load
120 echo "Client load failed on node $END_RUN_NODE"
122 echo "client $END_RUN_NODE load stdout and debug files :
123 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}
124 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}.debug"
130 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
131 echo "$0" >> $END_RUN_FILE
133 [ $rc -eq 0 ] || result=FAIL
135 log "Duraion: $DURATION
136 Server failover period: $SERVER_FAILOVER_PERIOD seconds
137 Exited after: $ELAPSED seconds
138 Number of failovers before exit:
140 Status: $result: rc=$rc"
142 # stop the vmstats on the OSTs
143 if [ "$VMSTAT" ]; then
144 do_nodes $(comma_list $(osts_nodes)) "test -f /tmp/vmstat.pid && \
145 { kill -s TERM \$(cat /tmp/vmstat.pid); rm -f /tmp/vmstat.pid; \
146 gzip -f9 $vmstatLOG-\$(hostname); }"
149 # make sure the client loads die
150 do_nodes $NODES_TO_USE "set -x; test -f $LOAD_PID_FILE && \
151 { kill -s TERM \$(cat $LOAD_PID_FILE) || true; }"
153 # and free up the pdshes that started them, if any are still around
154 if [ -n "$CLIENT_LOAD_PIDS" ]; then
155 kill $CLIENT_LOAD_PIDS || true
157 kill -9 $CLIENT_LOAD_PIDS || true
160 if [ $rc -ne 0 ]; then
161 print_logs $NODES_TO_USE
162 # we are interested in only on failed clients and servers
163 local failedclients=$(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 $LOAD_PID_FILE"; then
200 MINSLEEP=${MINSLEEP:-120}
201 REQFAIL_PERCENT=${REQFAIL_PERCENT:-3} # bug17839 comment 62
202 REQFAIL=${REQFAIL:-$(( DURATION / SERVER_FAILOVER_PERIOD * REQFAIL_PERCENT / 100))}
206 # This is used for FAIL_CLIENT only
208 while [ $ELAPSED -lt $DURATION -a ! -e $END_RUN_FILE ]; do
210 # In order to perform the
211 # expected number of failovers, we need to account the following :
212 # 1) the time that has elapsed during the client load checking
213 # 2) time takes for failover
215 it_time_start=$(date +%s)
217 FAIL_CLIENT=$(get_random_entry $NODES_TO_USE)
218 client_var=$(client_var_name $FAIL_CLIENT)_nums
220 # store the list of failed clients
221 # lists are comma separated
222 failed_clients=$(expand_list $failed_clients $FAIL_CLIENT)
224 SERVERFACET=$(get_random_entry $MDTS)
225 var=${SERVERFACET}_nums
227 # Check that our client loads are still running. If any have died,
228 # that means they have died outside of recovery, which is unacceptable.
230 log "==== Checking the clients loads BEFORE failover -- failure NOT OK \
231 ELAPSED=$ELAPSED DURATION=$DURATION PERIOD=$SERVER_FAILOVER_PERIOD"
233 if ! check_client_loads $NODES_TO_USE; then
237 log "FAIL CLIENT $FAIL_CLIENT ... "
238 shutdown_client $FAIL_CLIENT
240 log "Starting failover on $SERVERFACET"
242 facet_failover "$SERVERFACET" || exit 1
243 if ! wait_recovery_complete $SERVERFACET $((TIMEOUT * 10)); then
244 echo "$SERVERFACET recovery is not completed!"
248 boot_node $FAIL_CLIENT
249 echo "Reintegrating $FAIL_CLIENT"
250 zconf_mount $FAIL_CLIENT $MOUNT || exit $?
252 # Increment the number of failovers
255 val=$((${!client_var} + 1))
256 eval $client_var=$val
258 # load script on failed clients could create END_RUN_FILE
259 # We shuold remove it and ignore the failure if this
260 # file contains the failed client only.
261 # We can not use ERRORS_OK when start all loads at the start of this script
262 # because the application errors allowed for random failed client only, but
263 # not for all clients.
264 if [ -e $END_RUN_FILE ]; then
265 read END_RUN_NODE < $END_RUN_FILE
266 [[ $END_RUN_NODE = $FAIL_CLIENT ]] &&
267 rm -f $END_RUN_FILE || exit 13
270 restart_client_loads $FAIL_CLIENT $ERRORS_OK || exit $?
272 # Check that not failed clients loads are still running.
273 # No application failures should occur on clients that was not failed.
275 log "==== Checking the clients loads AFTER failed client reintegrated -- failure NOT OK"
276 if ! ERRORS_OK= check_client_loads $(exclude_items_from_list $NODES_TO_USE $FAIL_CLIENT); then
277 log "Client load failed. Exiting"
281 CURRENT_TS=$(date +%s)
282 ELAPSED=$((CURRENT_TS - START_TS))
284 sleep=$((SERVER_FAILOVER_PERIOD-(CURRENT_TS - it_time_start)))
286 # keep count the number of itterations when
287 # time spend to failover and two client loads check exceeded
288 # the value ( SERVER_FAILOVER_PERIOD - MINSLEEP )
289 if [ $sleep -lt $MINSLEEP ]; then
290 reqfail=$((reqfail +1))
291 log "WARNING: failover, client reintegration and check_client_loads time exceeded SERVER_FAILOVER_PERIOD - MINSLEEP !
292 Failed to load the filesystem with I/O for a minimum period of $MINSLEEP $reqfail times ( REQFAIL=$REQFAIL ).
293 This iteration, the load was only applied for sleep=$sleep seconds.
294 Probably the hardware is taking excessively long to boot.
295 Try to increase SERVER_FAILOVER_PERIOD (current is $SERVER_FAILOVER_PERIOD), bug 20918"
296 [ $reqfail -gt $REQFAIL ] && exit 6
299 log " Number of failovers:
300 $(numfailovers) and counting..."
302 if [ $((ELAPSED + sleep)) -gt $DURATION ]; then
306 if [ $sleep -gt 0 ]; then
307 echo "sleeping $sleep seconds ... "