7 # This test needs to be run on the client
10 export MULTIOP=${MULTIOP:-multiop}
11 LUSTRE=${LUSTRE:-$(cd $(dirname $0)/..; echo $PWD)}
14 . $LUSTRE/tests/test-framework.sh
16 . ${CONFIG:=$LUSTRE/tests/cfg/$NAME.sh}
18 CHECK_GRANT=${CHECK_GRANT:-"yes"}
19 GRANT_CHECK_LIST=${GRANT_CHECK_LIST:-""}
21 require_dsh_mds || exit 0
24 # bug number: 17466 18857 LU-1867 LU-1473
25 ALWAYS_EXCEPT="61d 33a 33b 89 62 $REPLAY_SINGLE_EXCEPT"
27 [ $(facet_fstype $SINGLEMDS) = "zfs" ] &&
28 # bug number for skipped test: LU-951
29 ALWAYS_EXCEPT="$ALWAYS_EXCEPT 73a"
31 # 63 min 7 min AT AT AT AT"
32 [ "$SLOW" = "no" ] && EXCEPT_SLOW="1 2 3 4 6 12 16 44a 44b 65 66 67 68"
34 [ $(facet_fstype $SINGLEMDS) = "zfs" ] &&
35 # bug number for skipped test: LU-3127
36 ALWAYS_EXCEPT="$ALWAYS_EXCEPT 73b"
40 check_and_setup_lustre
45 rm -rf $DIR/[df][0-9]* $DIR/f.$TESTSUITE.*
47 # LU-482 Avert LVM and VM inability to flush caches in pre .33 kernels
48 if [ $LINUX_VERSION_CODE -lt $(version_code 2.6.33) ]; then
50 do_facet $SINGLEMDS sync
53 test_0a() { # was test_0
55 replay_barrier $SINGLEMDS
59 run_test 0a "empty replay"
62 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
64 # this test attempts to trigger a race in the precreation code,
65 # and must run before any other objects are created on the filesystem
67 createmany -o $DIR/$tfile 20 || return 1
68 unlinkmany $DIR/$tfile 20 || return 2
70 run_test 0b "ensure object created after recover exists. (3284)"
73 replay_barrier $SINGLEMDS
76 facet_failover $SINGLEMDS
77 zconf_mount `hostname` $MOUNT || error "mount fails"
78 client_up || error "post-failover df failed"
79 # file shouldn't exist if replay-barrier works as expected
80 rm $DIR/$tfile && error "File exists and it shouldn't"
83 run_test 0c "check replay-barrier"
86 replay_barrier $SINGLEMDS
88 facet_failover $SINGLEMDS
89 zconf_mount `hostname` $MOUNT || error "mount fails"
90 client_up || error "post-failover df failed"
92 run_test 0d "expired recovery with no clients"
95 replay_barrier $SINGLEMDS
98 $CHECKSTAT -t file $DIR/$tfile || return 1
101 run_test 1 "simple create"
104 replay_barrier $SINGLEMDS
107 $CHECKSTAT -t file $DIR/$tfile || return 1
114 replay_barrier $SINGLEMDS
117 $CHECKSTAT -t file $DIR/$tfile || return 1
123 local file=$DIR/$tfile
124 replay_barrier $SINGLEMDS
126 openfile -f O_DIRECTORY $file
128 $CHECKSTAT -t file $file || return 2
131 run_test 3a "replay failed open(O_DIRECTORY)"
134 replay_barrier $SINGLEMDS
135 #define OBD_FAIL_MDS_OPEN_PACK | OBD_FAIL_ONCE
136 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000114"
138 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
140 $CHECKSTAT -t file $DIR/$tfile && return 2
143 run_test 3b "replay failed open -ENOMEM"
146 replay_barrier $SINGLEMDS
147 #define OBD_FAIL_MDS_ALLOC_OBDO | OBD_FAIL_ONCE
148 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000128"
150 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
153 $CHECKSTAT -t file $DIR/$tfile && return 2
156 run_test 3c "replay failed open -ENOMEM"
158 test_4a() { # was test_4
159 replay_barrier $SINGLEMDS
160 for i in `seq 10`; do
161 echo "tag-$i" > $DIR/$tfile-$i
164 for i in `seq 10`; do
165 grep -q "tag-$i" $DIR/$tfile-$i || error "$tfile-$i"
168 run_test 4a "|x| 10 open(O_CREAT)s"
171 replay_barrier $SINGLEMDS
174 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
176 run_test 4b "|x| rm 10 files"
178 # The idea is to get past the first block of precreated files on both
179 # osts, and then replay.
181 replay_barrier $SINGLEMDS
182 for i in `seq 220`; do
183 echo "tag-$i" > $DIR/$tfile-$i
186 for i in `seq 220`; do
187 grep -q "tag-$i" $DIR/$tfile-$i || error "$tfile-$i"
191 # waiting for commitment of removal
193 run_test 5 "|x| 220 open(O_CREAT)"
196 test_6a() { # was test_6
198 replay_barrier $SINGLEMDS
199 mcreate $DIR/$tdir/$tfile
201 $CHECKSTAT -t dir $DIR/$tdir || return 1
202 $CHECKSTAT -t file $DIR/$tdir/$tfile || return 2
204 # waiting for log process thread
206 run_test 6a "mkdir + contained create"
210 replay_barrier $SINGLEMDS
213 $CHECKSTAT -t dir $DIR/$tdir && return 1 || true
215 run_test 6b "|X| rmdir"
219 replay_barrier $SINGLEMDS
220 mcreate $DIR/$tdir/$tfile
222 $CHECKSTAT -t dir $DIR/$tdir || return 1
223 $CHECKSTAT -t file $DIR/$tdir/$tfile || return 2
226 run_test 7 "mkdir |X| contained create"
229 # make sure no side-effect from previous test.
231 replay_barrier $SINGLEMDS
232 multiop_bg_pause $DIR/$tfile mo_c || return 4
236 $CHECKSTAT -t file $DIR/$tfile || return 1
237 kill -USR1 $MULTIPID || return 2
238 wait $MULTIPID || return 3
241 run_test 8 "creat open |X| close"
244 replay_barrier $SINGLEMDS
246 local old_inum=`ls -i $DIR/$tfile | awk '{print $1}'`
248 local new_inum=`ls -i $DIR/$tfile | awk '{print $1}'`
250 echo " old_inum == $old_inum, new_inum == $new_inum"
251 if [ $old_inum -eq $new_inum ] ;
253 echo " old_inum and new_inum match"
255 echo "!!!! old_inum and new_inum NOT match"
260 run_test 9 "|X| create (same inum/gen)"
264 replay_barrier $SINGLEMDS
265 mv $DIR/$tfile $DIR/$tfile-2
268 $CHECKSTAT $DIR/$tfile && return 1
269 $CHECKSTAT $DIR/$tfile-2 ||return 2
273 run_test 10 "create |X| rename unlink"
277 echo "old" > $DIR/$tfile
278 mv $DIR/$tfile $DIR/$tfile-2
279 replay_barrier $SINGLEMDS
280 echo "new" > $DIR/$tfile
282 grep old $DIR/$tfile-2
284 grep new $DIR/$tfile || return 1
285 grep old $DIR/$tfile-2 || return 2
287 run_test 11 "create open write rename |X| create-old-name read"
291 multiop_bg_pause $DIR/$tfile o_tSc || return 3
294 replay_barrier $SINGLEMDS
296 wait $pid || return 1
299 [ -e $DIR/$tfile ] && return 2
302 run_test 12 "open, unlink |X| close"
305 # 1777 - replay open after committed chmod that would make
306 # a regular open a failure
309 multiop_bg_pause $DIR/$tfile O_wc || return 3
312 $CHECKSTAT -p 0 $DIR/$tfile
313 replay_barrier $SINGLEMDS
316 wait $pid || return 1
318 $CHECKSTAT -s 1 -p 0 $DIR/$tfile || return 2
319 rm $DIR/$tfile || return 4
322 run_test 13 "open chmod 0 |x| write close"
325 multiop_bg_pause $DIR/$tfile O_tSc || return 4
328 replay_barrier $SINGLEMDS
329 kill -USR1 $pid || return 1
330 wait $pid || return 2
333 [ -e $DIR/$tfile ] && return 3
336 run_test 14 "open(O_CREAT), unlink |X| close"
339 multiop_bg_pause $DIR/$tfile O_tSc || return 5
342 replay_barrier $SINGLEMDS
343 touch $DIR/g11 || return 1
345 wait $pid || return 2
348 [ -e $DIR/$tfile ] && return 3
349 touch $DIR/h11 || return 4
352 run_test 15 "open(O_CREAT), unlink |X| touch new, close"
356 replay_barrier $SINGLEMDS
359 mcreate $DIR/$tfile-2
361 [ -e $DIR/$tfile ] && return 1
362 [ -e $DIR/$tfile-2 ] || return 2
363 munlink $DIR/$tfile-2 || return 3
365 run_test 16 "|X| open(O_CREAT), unlink, touch new, unlink new"
368 replay_barrier $SINGLEMDS
369 multiop_bg_pause $DIR/$tfile O_c || return 4
372 kill -USR1 $pid || return 1
373 wait $pid || return 2
374 $CHECKSTAT -t file $DIR/$tfile || return 3
377 run_test 17 "|X| open(O_CREAT), |replay| close"
380 replay_barrier $SINGLEMDS
381 multiop_bg_pause $DIR/$tfile O_tSc || return 8
384 touch $DIR/$tfile-2 || return 1
385 echo "pid: $pid will close"
387 wait $pid || return 2
390 [ -e $DIR/$tfile ] && return 3
391 [ -e $DIR/$tfile-2 ] || return 4
392 # this touch frequently fails
393 touch $DIR/$tfile-3 || return 5
394 munlink $DIR/$tfile-2 || return 6
395 munlink $DIR/$tfile-3 || return 7
398 run_test 18 "|X| open(O_CREAT), unlink, touch new, close, touch, unlink"
400 # bug 1855 (a simpler form of test_11 above)
402 replay_barrier $SINGLEMDS
404 echo "old" > $DIR/$tfile
405 mv $DIR/$tfile $DIR/$tfile-2
406 grep old $DIR/$tfile-2
408 grep old $DIR/$tfile-2 || return 2
410 run_test 19 "|X| mcreate, open, write, rename "
412 test_20a() { # was test_20
413 replay_barrier $SINGLEMDS
414 multiop_bg_pause $DIR/$tfile O_tSc || return 3
420 wait $pid || return 1
421 [ -e $DIR/$tfile ] && return 2
424 run_test 20a "|X| open(O_CREAT), unlink, replay, close (test mds_cleanup_orphans)"
426 test_20b() { # bug 10480
427 local wait_timeout=$((TIMEOUT * 4))
431 BEFOREUSED=`df -P $DIR | tail -1 | awk '{ print $3 }'`
432 dd if=/dev/zero of=$DIR/$tfile bs=4k count=10000 &
434 while [ ! -e $DIR/$tfile ] ; do
435 usleep 60 # give dd a chance to start
438 $GETSTRIPE $DIR/$tfile || return 1
439 rm -f $DIR/$tfile || return 2 # make it an orphan
441 client_up || client_up || true # reconnect
443 do_facet $SINGLEMDS "lctl set_param -n osd*.*MDT*.force_sync 1"
445 fail $SINGLEMDS # start orphan recovery
446 wait_recovery_complete $SINGLEMDS || error "MDS recovery not done"
447 wait_delete_completed_mds $wait_timeout || return 3
449 AFTERUSED=$(df -P $DIR | tail -1 | awk '{ print $3 }')
450 log "before $BEFOREUSED, after $AFTERUSED"
451 (( $AFTERUSED > $BEFOREUSED + $(fs_log_size) )) &&
452 error "after $AFTERUSED > before $BEFOREUSED"
455 run_test 20b "write, unlink, eviction, replay, (test mds_cleanup_orphans)"
457 test_20c() { # bug 10480
458 multiop_bg_pause $DIR/$tfile Ow_c || return 1
464 client_up || client_up || true # reconnect
467 wait $pid || return 1
468 [ -s $DIR/$tfile ] || error "File was truncated"
472 run_test 20c "check that client eviction does not affect file content"
475 replay_barrier $SINGLEMDS
476 multiop_bg_pause $DIR/$tfile O_tSc || return 5
479 touch $DIR/g11 || return 1
483 wait $pid || return 2
484 [ -e $DIR/$tfile ] && return 3
485 touch $DIR/h11 || return 4
488 run_test 21 "|X| open(O_CREAT), unlink touch new, replay, close (test mds_cleanup_orphans)"
491 multiop_bg_pause $DIR/$tfile O_tSc || return 3
494 replay_barrier $SINGLEMDS
499 wait $pid || return 1
500 [ -e $DIR/$tfile ] && return 2
503 run_test 22 "open(O_CREAT), |X| unlink, replay, close (test mds_cleanup_orphans)"
506 multiop_bg_pause $DIR/$tfile O_tSc || return 5
509 replay_barrier $SINGLEMDS
511 touch $DIR/g11 || return 1
515 wait $pid || return 2
516 [ -e $DIR/$tfile ] && return 3
517 touch $DIR/h11 || return 4
520 run_test 23 "open(O_CREAT), |X| unlink touch new, replay, close (test mds_cleanup_orphans)"
523 multiop_bg_pause $DIR/$tfile O_tSc || return 3
526 replay_barrier $SINGLEMDS
530 wait $pid || return 1
531 [ -e $DIR/$tfile ] && return 2
534 run_test 24 "open(O_CREAT), replay, unlink, close (test mds_cleanup_orphans)"
537 multiop_bg_pause $DIR/$tfile O_tSc || return 3
541 replay_barrier $SINGLEMDS
544 wait $pid || return 1
545 [ -e $DIR/$tfile ] && return 2
548 run_test 25 "open(O_CREAT), unlink, replay, close (test mds_cleanup_orphans)"
551 replay_barrier $SINGLEMDS
552 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
554 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
559 wait $pid2 || return 1
563 wait $pid1 || return 2
564 [ -e $DIR/$tfile-1 ] && return 3
565 [ -e $DIR/$tfile-2 ] && return 4
568 run_test 26 "|X| open(O_CREAT), unlink two, close one, replay, close one (test mds_cleanup_orphans)"
571 replay_barrier $SINGLEMDS
572 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
574 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
581 wait $pid1 || return 1
583 wait $pid2 || return 2
584 [ -e $DIR/$tfile-1 ] && return 3
585 [ -e $DIR/$tfile-2 ] && return 4
588 run_test 27 "|X| open(O_CREAT), unlink two, replay, close two (test mds_cleanup_orphans)"
591 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
593 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
595 replay_barrier $SINGLEMDS
599 wait $pid2 || return 1
603 wait $pid1 || return 2
604 [ -e $DIR/$tfile-1 ] && return 3
605 [ -e $DIR/$tfile-2 ] && return 4
608 run_test 28 "open(O_CREAT), |X| unlink two, close one, replay, close one (test mds_cleanup_orphans)"
611 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
613 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
615 replay_barrier $SINGLEMDS
621 wait $pid1 || return 1
623 wait $pid2 || return 2
624 [ -e $DIR/$tfile-1 ] && return 3
625 [ -e $DIR/$tfile-2 ] && return 4
628 run_test 29 "open(O_CREAT), |X| unlink two, replay, close two (test mds_cleanup_orphans)"
631 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
633 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
638 replay_barrier $SINGLEMDS
641 wait $pid1 || return 1
643 wait $pid2 || return 2
644 [ -e $DIR/$tfile-1 ] && return 3
645 [ -e $DIR/$tfile-2 ] && return 4
648 run_test 30 "open(O_CREAT) two, unlink two, replay, close two (test mds_cleanup_orphans)"
651 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
653 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
657 replay_barrier $SINGLEMDS
661 wait $pid1 || return 1
663 wait $pid2 || return 2
664 [ -e $DIR/$tfile-1 ] && return 3
665 [ -e $DIR/$tfile-2 ] && return 4
668 run_test 31 "open(O_CREAT) two, unlink one, |X| unlink one, close two (test mds_cleanup_orphans)"
670 # tests for bug 2104; completion without crashing is success. The close is
671 # stale, but we always return 0 for close, so the app never sees it.
673 multiop_bg_pause $DIR/$tfile O_c || return 2
675 multiop_bg_pause $DIR/$tfile O_c || return 3
678 client_up || client_up || return 1
681 wait $pid1 || return 4
682 wait $pid2 || return 5
685 run_test 32 "close() notices client eviction; close() after client eviction"
688 createmany -o $DIR/$tfile-%d 10
689 replay_barrier_nosync $SINGLEMDS
690 fail_abort $SINGLEMDS
691 # recreate shouldn't fail
692 createmany -o $DIR/$tfile--%d 10 || return 1
696 run_test 33a "fid seq shouldn't be reused after abort recovery"
699 #define OBD_FAIL_SEQ_ALLOC 0x1311
700 do_facet $SINGLEMDS "lctl set_param fail_loc=0x1311"
702 createmany -o $DIR/$tfile-%d 10
703 replay_barrier_nosync $SINGLEMDS
704 fail_abort $SINGLEMDS
705 # recreate shouldn't fail
706 createmany -o $DIR/$tfile--%d 10 || return 1
710 run_test 33b "test fid seq allocation"
713 multiop_bg_pause $DIR/$tfile O_c || return 2
717 replay_barrier $SINGLEMDS
718 fail_abort $SINGLEMDS
720 wait $pid || return 3
721 [ -e $DIR/$tfile ] && return 1
725 run_test 34 "abort recovery before client does replay (test mds_cleanup_orphans)"
727 # bug 2278 - generate one orphan on OST, then destroy it during recovery from llog
731 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
732 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000119"
737 # give a chance to remove from MDS
738 fail_abort $SINGLEMDS
739 $CHECKSTAT -t file $DIR/$tfile && return 1 || true
741 run_test 35 "test recovery from llog for unlink op"
743 # b=2432 resent cancel after replay uses wrong cookie,
744 # so don't resend cancels
746 replay_barrier $SINGLEMDS
748 checkstat $DIR/$tfile
749 facet_failover $SINGLEMDS
751 if dmesg | grep "unknown lock cookie"; then
752 echo "cancel after replay failed"
756 run_test 36 "don't resend cancel"
759 # directory orphans can't be unlinked from PENDING directory
761 rmdir $DIR/$tfile 2>/dev/null
762 multiop_bg_pause $DIR/$tfile dD_c || return 2
766 replay_barrier $SINGLEMDS
767 # clear the dmesg buffer so we only see errors from this recovery
768 do_facet $SINGLEMDS dmesg -c >/dev/null
769 fail_abort $SINGLEMDS
771 do_facet $SINGLEMDS dmesg | grep "error .* unlinking .* from PENDING" &&
773 wait $pid || return 3
777 start_full_debug_logging
778 run_test 37 "abort recovery before client does replay (test mds_cleanup_orphans for directories)"
779 stop_full_debug_logging
782 createmany -o $DIR/$tfile-%d 800
783 unlinkmany $DIR/$tfile-%d 0 400
784 replay_barrier $SINGLEMDS
786 unlinkmany $DIR/$tfile-%d 400 400
788 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
790 run_test 38 "test recovery from unlink llog (test llog_gen_rec) "
792 test_39() { # bug 4176
793 createmany -o $DIR/$tfile-%d 800
794 replay_barrier $SINGLEMDS
795 unlinkmany $DIR/$tfile-%d 0 400
797 unlinkmany $DIR/$tfile-%d 400 400
799 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
801 run_test 39 "test recovery from unlink llog (test llog_gen_rec) "
804 lctl get_param -n osc.*.stats | awk -vwrites=0 '/ost_write/ { writes += $2 } END { print writes; }'
809 # always need connection to MDS to verify layout during IO. LU-2628.
810 lctl get_param mdc.*.connect_flags | grep -q layout_lock &&
811 skip "layout_lock needs MDS connection for IO" && return 0
813 $LCTL mark multiop $MOUNT/$tfile OS_c
814 multiop $MOUNT/$tfile OS_c &
816 writeme -s $MOUNT/${tfile}-2 &
819 facet_failover $SINGLEMDS
820 #define OBD_FAIL_MDS_CONNECT_NET 0x117
821 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000117"
823 stat1=`count_ost_writes`
825 stat2=`count_ost_writes`
826 echo "$stat1, $stat2"
827 if [ $stat1 -lt $stat2 ]; then
828 echo "writes continuing during recovery"
831 echo "writes not continuing during recovery, bug 2477"
834 echo "waiting for writeme $WRITE_PID"
838 echo "waiting for multiop $PID"
839 wait $PID || return 2
840 do_facet client munlink $MOUNT/$tfile || return 3
841 do_facet client munlink $MOUNT/${tfile}-2 || return 3
844 run_test 40 "cause recovery in ptlrpc, ensure IO continues"
848 # make sure that a read to one osc doesn't try to double-unlock its page just
849 # because another osc is invalid. trigger_group_io used to mistakenly return
850 # an error if any oscs were invalid even after having successfully put rpcs
851 # on valid oscs. This was fatal if the caller was ll_readpage who unlocked
852 # the page, guarnateeing that the unlock from the RPC completion would
853 # assert on trying to unlock the unlocked page.
855 [ $OSTCOUNT -lt 2 ] &&
856 skip_env "skipping test 41: we don't have a second OST to test with" &&
859 local f=$MOUNT/$tfile
860 # make sure the start of the file is ost1
861 $SETSTRIPE -S $((128 * 1024)) -i 0 $f
862 do_facet client dd if=/dev/zero of=$f bs=4k count=1 || return 3
864 # fail ost2 and read from ost1
865 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $ost2_svc)
866 local osc2dev=$(do_facet $SINGLEMDS "lctl get_param -n devices" | \
867 grep $mdtosc | awk '{print $1}')
868 [ -z "$osc2dev" ] && echo "OST: $ost2_svc" && lctl get_param -n devices &&
870 do_facet $SINGLEMDS $LCTL --device $osc2dev deactivate || return 1
871 do_facet client dd if=$f of=/dev/null bs=4k count=1 || return 3
872 do_facet $SINGLEMDS $LCTL --device $osc2dev activate || return 2
875 run_test 41 "read from a valid osc while other oscs are invalid"
877 # test MDS recovery after ost failure
879 blocks=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
880 createmany -o $DIR/$tfile-%d 800
882 unlinkmany $DIR/$tfile-%d 0 400
884 lctl set_param debug=-1
887 # osc is evicted, fs is smaller (but only with failout OSTs (bug 7287)
888 #blocks_after=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
889 #[ $blocks_after -lt $blocks ] || return 1
890 echo wait for MDS to timeout and recover
891 sleep $((TIMEOUT * 2))
893 unlinkmany $DIR/$tfile-%d 400 400
894 $CHECKSTAT -t file $DIR/$tfile-* && return 2 || true
896 run_test 42 "recovery after ost failure"
898 # timeout in MDS/OST recovery RPC will LBUG MDS
899 test_43() { # bug 2530
900 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
902 replay_barrier $SINGLEMDS
904 # OBD_FAIL_OST_CREATE_NET 0x204
905 do_facet ost1 "lctl set_param fail_loc=0x80000204"
908 do_facet ost1 "lctl set_param fail_loc=0"
912 run_test 43 "mds osc import failure during recovery; don't LBUG"
914 test_44a() { # was test_44
917 local mdcdev=$($LCTL dl |
918 awk "/${FSNAME}-MDT0000-mdc-/ {if (\$2 == \"UP\") {print \$1}}")
919 [ "$mdcdev" ] || return 2
920 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
921 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
923 # adaptive timeouts slow this way down
924 if at_is_enabled; then
925 at_max_saved=$(at_max_get mds)
929 for i in `seq 1 10`; do
930 echo "$i of 10 ($(date +%s))"
931 do_facet $SINGLEMDS \
932 "lctl get_param -n md[ts].*.mdt.timeouts | grep service"
933 #define OBD_FAIL_TGT_CONN_RACE 0x701
934 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000701"
935 # lctl below may fail, it is valid case
936 $LCTL --device $mdcdev recover
939 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
940 [ $at_max_saved -ne 0 ] && at_max_set $at_max_saved mds
943 run_test 44a "race in target handle connect"
946 local mdcdev=$($LCTL dl |
947 awk "/${FSNAME}-MDT0000-mdc-/ {if (\$2 == \"UP\") {print \$1}}")
948 [ "$mdcdev" ] || return 2
949 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
950 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
952 for i in `seq 1 10`; do
953 echo "$i of 10 ($(date +%s))"
954 do_facet $SINGLEMDS \
955 "lctl get_param -n md[ts].*.mdt.timeouts | grep service"
956 #define OBD_FAIL_TGT_DELAY_RECONNECT 0x704
957 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000704"
958 # lctl below may fail, it is valid case
959 $LCTL --device $mdcdev recover
962 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
965 run_test 44b "race in target handle connect"
968 replay_barrier $SINGLEMDS
969 createmany -m $DIR/$tfile-%d 100 || error "failed to create directories"
970 #define OBD_FAIL_TGT_RCVG_FLAG 0x712
971 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000712"
972 fail_abort $SINGLEMDS
973 unlinkmany $DIR/$tfile-%d 100 && error "unliked after fail abort"
975 unlinkmany $DIR/$tfile-%d 100 && error "unliked after fail"
978 run_test 44c "race in target handle connect"
980 # Handle failed close
982 local mdcdev=$($LCTL get_param -n devices |
983 awk "/ ${FSNAME}-MDT0000-mdc-/ {print \$1}")
984 [ "$mdcdev" ] || return 2
985 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
986 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
988 $LCTL --device $mdcdev recover || return 6
990 multiop_bg_pause $DIR/$tfile O_c || return 1
993 # This will cause the CLOSE to fail before even
994 # allocating a reply buffer
995 $LCTL --device $mdcdev deactivate || return 4
999 wait $pid || return 1
1001 $LCTL --device $mdcdev activate || return 5
1004 $CHECKSTAT -t file $DIR/$tfile || return 2
1007 run_test 45 "Handle failed close"
1011 drop_reply "touch $DIR/$tfile"
1013 # ironically, the previous test, 45, will cause a real forced close,
1014 # so just look for one for this test
1015 dmesg | grep -i "force closing client file handle for $tfile" && return 1
1018 run_test 46 "Don't leak file handle after open resend (3325)"
1020 test_47() { # bug 2824
1021 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1023 # create some files to make sure precreate has been done on all
1024 # OSTs. (just in case this test is run independently)
1025 createmany -o $DIR/$tfile 20 || return 1
1027 # OBD_FAIL_OST_CREATE_NET 0x204
1029 do_facet ost1 "lctl set_param fail_loc=0x80000204"
1030 client_up || return 2
1032 # let the MDS discover the OST failure, attempt to recover, fail
1033 # and recover again.
1034 sleep $((3 * TIMEOUT))
1036 # Without 2824, this createmany would hang
1037 createmany -o $DIR/$tfile 20 || return 3
1038 unlinkmany $DIR/$tfile 20 || return 4
1040 do_facet ost1 "lctl set_param fail_loc=0"
1043 run_test 47 "MDS->OSC failure during precreate cleanup (2824)"
1046 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1047 [ "$OSTCOUNT" -lt "2" ] && skip_env "$OSTCOUNT < 2 OSTs -- skipping" && return
1049 replay_barrier $SINGLEMDS
1050 createmany -o $DIR/$tfile 20 || return 1
1051 # OBD_FAIL_OST_EROFS 0x216
1052 facet_failover $SINGLEMDS
1053 do_facet ost1 "lctl set_param fail_loc=0x80000216"
1054 client_up || return 2
1056 createmany -o $DIR/$tfile 20 20 || return 2
1057 unlinkmany $DIR/$tfile 40 || return 3
1060 run_test 48 "MDS->OSC failure during precreate cleanup (2824)"
1063 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $ost1_svc)
1064 local oscdev=$(do_facet $SINGLEMDS "lctl get_param -n devices" | \
1065 grep $mdtosc | awk '{print $1}')
1066 [ "$oscdev" ] || return 1
1067 do_facet $SINGLEMDS $LCTL --device $oscdev recover || return 2
1068 do_facet $SINGLEMDS $LCTL --device $oscdev recover || return 3
1069 # give the mds_lov_sync threads a chance to run
1072 run_test 50 "Double OSC recovery, don't LASSERT (3812)"
1074 # b3764 timed out lock replay
1077 cancel_lru_locks mdc
1079 multiop $DIR/$tfile s || return 1
1080 replay_barrier $SINGLEMDS
1081 #define OBD_FAIL_LDLM_REPLY 0x30c
1082 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000030c"
1083 fail $SINGLEMDS || return 2
1084 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1086 $CHECKSTAT -t file $DIR/$tfile-* && return 3 || true
1088 run_test 52 "time out lock replay (3764)"
1090 # bug 3462 - simultaneous MDC requests
1092 cancel_lru_locks mdc # cleanup locks from former test cases
1093 mkdir -p $DIR/${tdir}-1
1094 mkdir -p $DIR/${tdir}-2
1095 multiop $DIR/${tdir}-1/f O_c &
1097 # give multiop a change to open
1100 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1101 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1102 kill -USR1 $close_pid
1103 cancel_lru_locks mdc # force the close
1104 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1106 mcreate $DIR/${tdir}-2/f || return 1
1108 # close should still be here
1109 [ -d /proc/$close_pid ] || return 2
1111 replay_barrier_nodf $SINGLEMDS
1113 wait $close_pid || return 3
1115 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1116 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1117 rm -rf $DIR/${tdir}-*
1119 run_test 53a "|X| close request while two MDC requests in flight"
1122 cancel_lru_locks mdc # cleanup locks from former test cases
1123 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1125 mkdir -p $DIR/${tdir}-1
1126 mkdir -p $DIR/${tdir}-2
1127 multiop_bg_pause $DIR/${tdir}-1/f O_c || return 6
1130 #define OBD_FAIL_MDS_REINT_NET 0x107
1131 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1132 mcreate $DIR/${tdir}-2/f &
1136 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1137 kill -USR1 $close_pid
1138 cancel_lru_locks mdc # force the close
1139 wait $close_pid || return 1
1140 # open should still be here
1141 [ -d /proc/$open_pid ] || return 2
1143 replay_barrier_nodf $SINGLEMDS
1145 wait $open_pid || return 3
1147 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1148 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1149 rm -rf $DIR/${tdir}-*
1151 run_test 53b "|X| open request while two MDC requests in flight"
1154 cancel_lru_locks mdc # cleanup locks from former test cases
1155 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1157 mkdir -p $DIR/${tdir}-1
1158 mkdir -p $DIR/${tdir}-2
1159 multiop $DIR/${tdir}-1/f O_c &
1162 #define OBD_FAIL_MDS_REINT_NET 0x107
1163 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1164 mcreate $DIR/${tdir}-2/f &
1168 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1169 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1170 kill -USR1 $close_pid
1171 cancel_lru_locks mdc # force the close
1173 #bz20647: make sure all pids are exists before failover
1174 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1175 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1176 replay_barrier_nodf $SINGLEMDS
1177 fail_nodf $SINGLEMDS
1178 wait $open_pid || return 1
1180 # close should be gone
1181 [ -d /proc/$close_pid ] && return 2
1182 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1184 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1185 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1186 rm -rf $DIR/${tdir}-*
1188 run_test 53c "|X| open request and close request while two MDC requests in flight"
1191 cancel_lru_locks mdc # cleanup locks from former test cases
1192 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1194 mkdir -p $DIR/${tdir}-1
1195 mkdir -p $DIR/${tdir}-2
1196 multiop $DIR/${tdir}-1/f O_c &
1198 # give multiop a chance to open
1201 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1202 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1203 kill -USR1 $close_pid
1204 cancel_lru_locks mdc # force the close
1205 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1206 mcreate $DIR/${tdir}-2/f || return 1
1208 # close should still be here
1209 [ -d /proc/$close_pid ] || return 2
1211 wait $close_pid || return 3
1213 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1214 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1215 rm -rf $DIR/${tdir}-*
1217 run_test 53d "|X| close reply while two MDC requests in flight"
1220 cancel_lru_locks mdc # cleanup locks from former test cases
1221 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1223 mkdir -p $DIR/${tdir}-1
1224 mkdir -p $DIR/${tdir}-2
1225 multiop $DIR/${tdir}-1/f O_c &
1228 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1229 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1230 mcreate $DIR/${tdir}-2/f &
1234 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1235 kill -USR1 $close_pid
1236 cancel_lru_locks mdc # force the close
1237 wait $close_pid || return 1
1238 # open should still be here
1239 [ -d /proc/$open_pid ] || return 2
1241 replay_barrier_nodf $SINGLEMDS
1243 wait $open_pid || return 3
1245 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1246 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1247 rm -rf $DIR/${tdir}-*
1249 run_test 53e "|X| open reply while two MDC requests in flight"
1252 cancel_lru_locks mdc # cleanup locks from former test cases
1253 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1255 mkdir -p $DIR/${tdir}-1
1256 mkdir -p $DIR/${tdir}-2
1257 multiop $DIR/${tdir}-1/f O_c &
1260 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1261 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1262 mcreate $DIR/${tdir}-2/f &
1266 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1267 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1268 kill -USR1 $close_pid
1269 cancel_lru_locks mdc # force the close
1271 #bz20647: make sure all pids are exists before failover
1272 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1273 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1274 replay_barrier_nodf $SINGLEMDS
1275 fail_nodf $SINGLEMDS
1276 wait $open_pid || return 1
1278 # close should be gone
1279 [ -d /proc/$close_pid ] && return 2
1280 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1282 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1283 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1284 rm -rf $DIR/${tdir}-*
1286 run_test 53f "|X| open reply and close reply while two MDC requests in flight"
1289 cancel_lru_locks mdc # cleanup locks from former test cases
1290 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1292 mkdir -p $DIR/${tdir}-1
1293 mkdir -p $DIR/${tdir}-2
1294 multiop $DIR/${tdir}-1/f O_c &
1297 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1298 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1299 mcreate $DIR/${tdir}-2/f &
1303 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1304 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1305 kill -USR1 $close_pid
1306 cancel_lru_locks mdc # force the close
1307 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1309 #bz20647: make sure all pids are exists before failover
1310 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1311 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1312 replay_barrier_nodf $SINGLEMDS
1313 fail_nodf $SINGLEMDS
1314 wait $open_pid || return 1
1316 # close should be gone
1317 [ -d /proc/$close_pid ] && return 2
1319 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1320 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1321 rm -rf $DIR/${tdir}-*
1323 run_test 53g "|X| drop open reply and close request while close and open are both in flight"
1326 cancel_lru_locks mdc # cleanup locks from former test cases
1327 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1329 mkdir -p $DIR/${tdir}-1
1330 mkdir -p $DIR/${tdir}-2
1331 multiop $DIR/${tdir}-1/f O_c &
1334 #define OBD_FAIL_MDS_REINT_NET 0x107
1335 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1336 mcreate $DIR/${tdir}-2/f &
1340 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1341 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1342 kill -USR1 $close_pid
1343 cancel_lru_locks mdc # force the close
1346 #bz20647: make sure all pids are exists before failover
1347 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1348 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1349 replay_barrier_nodf $SINGLEMDS
1350 fail_nodf $SINGLEMDS
1351 wait $open_pid || return 1
1353 # close should be gone
1354 [ -d /proc/$close_pid ] && return 2
1355 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1357 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1358 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1359 rm -rf $DIR/${tdir}-*
1361 run_test 53h "|X| open request and close reply while two MDC requests in flight"
1363 #b_cray 54 "|X| open request and close reply while two MDC requests in flight"
1365 #b3761 ASSERTION(hash != 0) failed
1367 # OBD_FAIL_MDS_OPEN_CREATE | OBD_FAIL_ONCE
1368 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012b"
1370 # give touch a chance to run
1372 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1376 run_test 55 "let MDS_CHECK_RESENT return the original return code instead of 0"
1378 #b3440 ASSERTION(rec->ur_fid2->id) failed
1380 ln -s foo $DIR/$tfile
1381 replay_barrier $SINGLEMDS
1382 #drop_reply "cat $DIR/$tfile"
1386 run_test 56 "don't replay a symlink open request (3440)"
1388 #recovery one mds-ost setattr from llog
1390 #define OBD_FAIL_MDS_OST_SETATTR 0x12c
1391 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012c"
1393 replay_barrier $SINGLEMDS
1396 $CHECKSTAT -t file $DIR/$tfile || return 1
1397 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1400 run_test 57 "test recovery from llog for setattr op"
1403 zconf_umount `hostname` $MOUNT2
1407 #recovery many mds-ost setattr from llog
1410 #define OBD_FAIL_MDS_OST_SETATTR 0x12c
1411 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012c"
1412 createmany -o $DIR/$tdir/$tfile-%d 2500
1413 replay_barrier $SINGLEMDS
1416 $CHECKSTAT -t file $DIR/$tdir/$tfile-* >/dev/null || return 1
1417 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1418 unlinkmany $DIR/$tdir/$tfile-%d 2500
1421 run_test 58a "test recovery from llog for setattr op (test llog_gen_rec)"
1427 trap cleanup_58 EXIT
1429 large_xattr_enabled &&
1430 orig="$(generate_string $(max_xattr_size))" || orig="bar"
1432 mount_client $MOUNT2
1434 touch $DIR/$tdir/$tfile
1435 replay_barrier $SINGLEMDS
1436 setfattr -n trusted.foo -v $orig $DIR/$tdir/$tfile
1438 new=$(get_xattr_value trusted.foo $MOUNT2/$tdir/$tfile)
1439 [[ "$new" = "$orig" ]] || return 1
1440 rm -f $DIR/$tdir/$tfile
1444 run_test 58b "test replay of setxattr op"
1446 test_58c() { # bug 16570
1451 trap cleanup_58 EXIT
1453 if large_xattr_enabled; then
1454 local xattr_size=$(max_xattr_size)
1455 orig="$(generate_string $((xattr_size / 2)))"
1456 orig1="$(generate_string $xattr_size)"
1462 mount_client $MOUNT2
1464 touch $DIR/$tdir/$tfile
1465 drop_request "setfattr -n trusted.foo -v $orig $DIR/$tdir/$tfile" ||
1467 new=$(get_xattr_value trusted.foo $MOUNT2/$tdir/$tfile)
1468 [[ "$new" = "$orig" ]] || return 2
1469 drop_reint_reply "setfattr -n trusted.foo1 -v $orig1 $DIR/$tdir/$tfile" ||
1471 new=$(get_xattr_value trusted.foo1 $MOUNT2/$tdir/$tfile)
1472 [[ "$new" = "$orig1" ]] || return 4
1473 rm -f $DIR/$tdir/$tfile
1477 run_test 58c "resend/reconstruct setxattr op"
1479 # log_commit_thread vs filter_destroy race used to lead to import use after free
1482 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1485 createmany -o $DIR/$tdir/$tfile-%d 200
1487 unlinkmany $DIR/$tdir/$tfile-%d 200
1488 #define OBD_FAIL_PTLRPC_DELAY_RECOV 0x507
1489 do_facet ost1 "lctl set_param fail_loc=0x507"
1492 do_facet ost1 "lctl set_param fail_loc=0x0"
1496 run_test 59 "test log_commit_thread vs filter_destroy race"
1498 # race between add unlink llog vs cat log init in post_recovery (only for b1_6)
1499 # bug 12086: should no oops and No ctxt error for this test
1502 createmany -o $DIR/$tdir/$tfile-%d 200
1503 replay_barrier $SINGLEMDS
1504 unlinkmany $DIR/$tdir/$tfile-%d 0 100
1506 unlinkmany $DIR/$tdir/$tfile-%d 100 100
1507 local no_ctxt=`dmesg | grep "No ctxt"`
1508 [ -z "$no_ctxt" ] || error "ctxt is not initialized in recovery"
1510 run_test 60 "test llog post recovery init vs llog unlink"
1512 #test race llog recovery thread vs llog cleanup
1513 test_61a() { # was test_61
1514 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1517 createmany -o $DIR/$tdir/$tfile-%d 800
1519 # OBD_FAIL_OST_LLOG_RECOVERY_TIMEOUT 0x221
1520 unlinkmany $DIR/$tdir/$tfile-%d 800
1521 set_nodes_failloc "$(osts_nodes)" 0x80000221
1526 set_nodes_failloc "$(osts_nodes)" 0x0
1528 $CHECKSTAT -t file $DIR/$tdir/$tfile-* && return 1
1531 run_test 61a "test race llog recovery vs llog cleanup"
1533 #test race mds llog sync vs llog cleanup
1535 # OBD_FAIL_MDS_LLOG_SYNC_TIMEOUT 0x13a
1536 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013a"
1537 facet_failover $SINGLEMDS
1540 do_facet client dd if=/dev/zero of=$DIR/$tfile bs=4k count=1 || return 1
1542 run_test 61b "test race mds llog sync vs llog cleanup"
1544 #test race cancel cookie cb vs llog cleanup
1546 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1548 # OBD_FAIL_OST_CANCEL_COOKIE_TIMEOUT 0x222
1550 set_nodes_failloc "$(osts_nodes)" 0x80000222
1554 set_nodes_failloc "$(osts_nodes)" 0x0
1556 run_test 61c "test race mds llog sync vs llog cleanup"
1558 test_61d() { # bug 16002 # bug 17466 # bug 22137
1559 # OBD_FAIL_OBD_LLOG_SETUP 0x605
1561 do_facet mgs "lctl set_param fail_loc=0x80000605"
1562 start mgs $MGSDEV $MGS_MOUNT_OPTS && error "mgs start should have failed"
1563 do_facet mgs "lctl set_param fail_loc=0"
1564 start mgs $MGSDEV $MGS_MOUNT_OPTS || error "cannot restart mgs"
1566 run_test 61d "error in llog_setup should cleanup the llog context correctly"
1568 test_62() { # Bug 15756 - don't mis-drop resent replay
1570 replay_barrier $SINGLEMDS
1571 createmany -o $DIR/$tdir/$tfile- 25
1572 #define OBD_FAIL_TGT_REPLAY_DROP 0x707
1573 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000707"
1575 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1576 unlinkmany $DIR/$tdir/$tfile- 25 || return 2
1579 run_test 62 "don't mis-drop resent replay"
1581 #Adaptive Timeouts (bug 3055)
1589 echo "Cleaning up AT ..."
1590 if [ -n "$ATOLDBASE" ]; then
1591 local at_history=$($LCTL get_param -n at_history)
1592 do_facet $SINGLEMDS "lctl set_param at_history=$at_history" || true
1593 do_facet ost1 "lctl set_param at_history=$at_history" || true
1596 if [ $AT_MAX_SET -ne 0 ]; then
1597 for facet in mds client ost; do
1598 var=AT_MAX_SAVE_${facet}
1599 echo restore AT on $facet to saved value ${!var}
1600 at_max_set ${!var} $facet
1601 at_new=$(at_max_get $facet)
1602 echo Restored AT value on $facet $at_new
1603 [ $at_new -eq ${!var} ] || \
1604 error "$facet : AT value was not restored SAVED ${!var} NEW $at_new"
1611 local at_max_new=600
1613 # Save at_max original values
1615 if [ $AT_MAX_SET -eq 0 ]; then
1616 # Suppose that all osts have the same at_max
1617 for facet in mds client ost; do
1618 eval AT_MAX_SAVE_${facet}=$(at_max_get $facet)
1622 for facet in mds client ost; do
1623 at_max=$(at_max_get $facet)
1624 if [ $at_max -ne $at_max_new ]; then
1625 echo "AT value on $facet is $at_max, set it by force temporarily to $at_max_new"
1626 at_max_set $at_max_new $facet
1631 if [ -z "$ATOLDBASE" ]; then
1632 ATOLDBASE=$(do_facet $SINGLEMDS "lctl get_param -n at_history")
1633 # speed up the timebase so we can check decreasing AT
1634 do_facet $SINGLEMDS "lctl set_param at_history=8" || true
1635 do_facet ost1 "lctl set_param at_history=8" || true
1637 # sleep for a while to cool down, should be > 8s and also allow
1638 # at least one ping to be sent. simply use TIMEOUT to be safe.
1643 test_65a() #bug 3055
1645 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1647 at_start || return 0
1648 $LCTL dk > /dev/null
1650 $LCTL set_param debug="other"
1651 # Slow down a request to the current service time, this is critical
1652 # because previous tests may have caused this value to increase.
1653 REQ_DELAY=`lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts |
1654 awk '/portal 12/ {print $5}'`
1655 REQ_DELAY=$((${REQ_DELAY} + ${REQ_DELAY} / 4 + 5))
1657 do_facet $SINGLEMDS lctl set_param fail_val=$((${REQ_DELAY} * 1000))
1658 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1659 do_facet $SINGLEMDS $LCTL set_param fail_loc=0x8000050a
1660 createmany -o $DIR/$tfile 10 > /dev/null
1661 unlinkmany $DIR/$tfile 10 > /dev/null
1662 # check for log message
1663 $LCTL dk | grep "Early reply #" || error "No early reply"
1665 # client should show REQ_DELAY estimates
1666 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep portal
1668 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep portal
1670 run_test 65a "AT: verify early replies"
1672 test_65b() #bug 3055
1674 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1676 at_start || return 0
1679 $LCTL set_param debug="other trace"
1680 $LCTL dk > /dev/null
1681 # Slow down a request to the current service time, this is critical
1682 # because previous tests may have caused this value to increase.
1683 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tfile
1684 multiop $DIR/$tfile Ow1yc
1685 REQ_DELAY=`lctl get_param -n osc.${FSNAME}-OST0000-osc-*.timeouts |
1686 awk '/portal 6/ {print $5}'`
1687 REQ_DELAY=$((${REQ_DELAY} + ${REQ_DELAY} / 4 + 5))
1689 do_facet ost1 lctl set_param fail_val=${REQ_DELAY}
1690 #define OBD_FAIL_OST_BRW_PAUSE_PACK 0x224
1691 do_facet ost1 $LCTL set_param fail_loc=0x224
1694 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tfile
1695 # force some real bulk transfer
1696 multiop $DIR/$tfile oO_CREAT:O_RDWR:O_SYNC:w4096c
1698 do_facet ost1 $LCTL set_param fail_loc=0
1699 # check for log message
1700 $LCTL dk | grep "Early reply #" || error "No early reply"
1702 # client should show REQ_DELAY estimates
1703 lctl get_param -n osc.${FSNAME}-OST0000-osc-*.timeouts | grep portal
1705 run_test 65b "AT: verify early replies on packed reply / bulk"
1707 test_66a() #bug 3055
1709 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1711 at_start || return 0
1712 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1713 # adjust 5s at a time so no early reply is sent (within deadline)
1714 do_facet $SINGLEMDS "$LCTL set_param fail_val=5000"
1715 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1716 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0x8000050a"
1717 createmany -o $DIR/$tfile 20 > /dev/null
1718 unlinkmany $DIR/$tfile 20 > /dev/null
1719 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1720 do_facet $SINGLEMDS "$LCTL set_param fail_val=10000"
1721 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0x8000050a"
1722 createmany -o $DIR/$tfile 20 > /dev/null
1723 unlinkmany $DIR/$tfile 20 > /dev/null
1724 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1725 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0"
1727 createmany -o $DIR/$tfile 20 > /dev/null
1728 unlinkmany $DIR/$tfile 20 > /dev/null
1729 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1730 CUR=$(lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | awk '/portal 12/ {print $5}')
1731 WORST=$(lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | awk '/portal 12/ {print $7}')
1732 echo "Current MDT timeout $CUR, worst $WORST"
1733 [ $CUR -lt $WORST ] || error "Current $CUR should be less than worst $WORST"
1735 run_test 66a "AT: verify MDT service time adjusts with no early replies"
1737 test_66b() #bug 3055
1739 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1741 at_start || return 0
1742 ORIG=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1743 awk '/network/ {print $4}')
1744 $LCTL set_param fail_val=$(($ORIG + 5))
1745 #define OBD_FAIL_PTLRPC_PAUSE_REP 0x50c
1746 $LCTL set_param fail_loc=0x50c
1747 ls $DIR/$tfile > /dev/null 2>&1
1748 $LCTL set_param fail_loc=0
1749 CUR=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1750 awk '/network/ {print $4}')
1751 WORST=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1752 awk '/network/ {print $6}')
1753 echo "network timeout orig $ORIG, cur $CUR, worst $WORST"
1754 [ $WORST -gt $ORIG ] ||
1755 error "Worst $WORST should be worse than orig $ORIG"
1757 run_test 66b "AT: verify net latency adjusts"
1759 test_67a() #bug 3055
1761 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1763 at_start || return 0
1764 CONN1=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1765 # sleeping threads may drive values above this
1766 do_facet ost1 "$LCTL set_param fail_val=400"
1767 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1768 do_facet ost1 "$LCTL set_param fail_loc=0x50a"
1769 createmany -o $DIR/$tfile 20 > /dev/null
1770 unlinkmany $DIR/$tfile 20 > /dev/null
1771 do_facet ost1 "$LCTL set_param fail_loc=0"
1772 CONN2=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1773 ATTEMPTS=$(($CONN2 - $CONN1))
1774 echo "$ATTEMPTS osc reconnect attempts on gradual slow"
1775 [ $ATTEMPTS -gt 0 ] && error_ignore 13721 "AT should have prevented reconnect"
1778 run_test 67a "AT: verify slow request processing doesn't induce reconnects"
1780 test_67b() #bug 3055
1782 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1784 at_start || return 0
1785 CONN1=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1787 # exhaust precreations on ost1
1788 local OST=$(ostname_from_index 0)
1789 local mdtosc=$(get_mdtosc_proc_path mds $OST)
1790 local last_id=$(do_facet $SINGLEMDS lctl get_param -n \
1791 osc.$mdtosc.prealloc_last_id)
1792 local next_id=$(do_facet $SINGLEMDS lctl get_param -n \
1793 osc.$mdtosc.prealloc_next_id)
1795 mkdir -p $DIR/$tdir/${OST}
1796 $SETSTRIPE -i 0 -c 1 $DIR/$tdir/${OST} || error "$SETSTRIPE"
1797 echo "Creating to objid $last_id on ost $OST..."
1798 #define OBD_FAIL_OST_PAUSE_CREATE 0x223
1799 do_facet ost1 "$LCTL set_param fail_val=20000"
1800 do_facet ost1 "$LCTL set_param fail_loc=0x80000223"
1801 createmany -o $DIR/$tdir/${OST}/f $next_id $((last_id - next_id + 2))
1804 do_facet ost1 "lctl get_param -n ost.OSS.ost_create.timeouts"
1806 CONN2=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1807 ATTEMPTS=$(($CONN2 - $CONN1))
1808 echo "$ATTEMPTS osc reconnect attempts on instant slow"
1809 # do it again; should not timeout
1810 do_facet ost1 "$LCTL set_param fail_loc=0x80000223"
1811 cp /etc/profile $DIR/$tfile || error "cp failed"
1812 do_facet ost1 "$LCTL set_param fail_loc=0"
1814 do_facet ost1 "lctl get_param -n ost.OSS.ost_create.timeouts"
1815 CONN3=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1816 ATTEMPTS=$(($CONN3 - $CONN2))
1817 echo "$ATTEMPTS osc reconnect attempts on 2nd slow"
1818 [ $ATTEMPTS -gt 0 ] && error "AT should have prevented reconnect"
1821 run_test 67b "AT: verify instant slowdown doesn't induce reconnects"
1823 test_68 () #bug 13813
1825 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1827 at_start || return 0
1828 local ldlm_enqueue_min=$(find /sys -name ldlm_enqueue_min)
1829 [ -z "$ldlm_enqueue_min" ] && skip "missing /sys/.../ldlm_enqueue_min" && return 0
1830 local ldlm_enqueue_min_r=$(do_facet ost1 "find /sys -name ldlm_enqueue_min")
1831 [ -z "$ldlm_enqueue_min_r" ] && skip "missing /sys/.../ldlm_enqueue_min in the ost1" && return 0
1832 local ENQ_MIN=$(cat $ldlm_enqueue_min)
1833 local ENQ_MIN_R=$(do_facet ost1 "cat $ldlm_enqueue_min_r")
1834 echo $TIMEOUT >> $ldlm_enqueue_min
1835 do_facet ost1 "echo $TIMEOUT >> $ldlm_enqueue_min_r"
1839 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tdir
1840 #define OBD_FAIL_LDLM_PAUSE_CANCEL 0x312
1841 $LCTL set_param fail_val=$(($TIMEOUT - 1))
1842 $LCTL set_param fail_loc=0x80000312
1843 cp /etc/profile $DIR/$tdir/${tfile}_1 || error "1st cp failed $?"
1844 $LCTL set_param fail_val=$((TIMEOUT * 5 / 4))
1845 $LCTL set_param fail_loc=0x80000312
1846 cp /etc/profile $DIR/$tdir/${tfile}_2 || error "2nd cp failed $?"
1847 $LCTL set_param fail_loc=0
1849 echo $ENQ_MIN >> $ldlm_enqueue_min
1850 do_facet ost1 "echo $ENQ_MIN_R >> $ldlm_enqueue_min_r"
1854 run_test 68 "AT: verify slowing locks"
1857 # end of AT tests includes above lines
1860 # start multi-client tests
1862 [ -z "$CLIENTS" ] && \
1863 { skip "Need two or more clients." && return; }
1864 [ $CLIENTCOUNT -lt 2 ] && \
1865 { skip "Need two or more clients, have $CLIENTCOUNT" && return; }
1867 echo "mount clients $CLIENTS ..."
1868 zconf_mount_clients $CLIENTS $MOUNT
1870 local clients=${CLIENTS//,/ }
1871 echo "Write/read files on $DIR ; clients $CLIENTS ... "
1872 for CLIENT in $clients; do
1873 do_node $CLIENT dd bs=1M count=10 if=/dev/zero \
1874 of=$DIR/${tfile}_${CLIENT} 2>/dev/null || \
1875 error "dd failed on $CLIENT"
1878 local prev_client=$(echo $clients | sed 's/^.* \(.\+\)$/\1/')
1879 for C in ${CLIENTS//,/ }; do
1880 do_node $prev_client dd if=$DIR/${tfile}_${C} of=/dev/null 2>/dev/null || \
1881 error "dd if=$DIR/${tfile}_${C} failed on $prev_client"
1887 run_test 70a "check multi client t-f"
1889 check_for_process () {
1894 killall_process $clients "$prog" -0
1897 killall_process () {
1898 local clients=${1:-$(hostname)}
1903 do_nodes $clients "killall $signal $name"
1907 local clients=${CLIENTS:-$HOSTNAME}
1909 zconf_mount_clients $clients $MOUNT
1912 [ "$SLOW" = "no" ] && duration=120
1913 # set duration to 900 because it takes some time to boot node
1914 [ "$FAILURE_MODE" = HARD ] && duration=900
1917 local start_ts=$(date +%s)
1918 local cmd="rundbench 1 -t $duration"
1920 do_nodesv $clients "set -x; MISSING_DBENCH_OK=$MISSING_DBENCH_OK \
1921 PATH=\$PATH:$LUSTRE/utils:$LUSTRE/tests/:$DBENCH_LIB \
1922 DBENCH_LIB=$DBENCH_LIB TESTSUITE=$TESTSUITE TESTNAME=$TESTNAME \
1923 MOUNT=$MOUNT DIR=$DIR/$tdir/\\\$(hostname) LCTL=$LCTL $cmd" &
1926 #LU-1897 wait for all dbench copies to start
1927 while ! check_for_process $clients dbench; do
1928 elapsed=$(($(date +%s) - start_ts))
1929 if [ $elapsed -gt $duration ]; then
1930 killall_process $clients dbench
1931 error "dbench failed to start on $clients!"
1936 log "Started rundbench load pid=$pid ..."
1938 elapsed=$(($(date +%s) - start_ts))
1939 local num_failovers=0
1940 while [ $elapsed -lt $duration ]; do
1941 if ! check_for_process $clients dbench; then
1942 error_noexit "dbench stopped on some of $clients!"
1943 killall_process $clients dbench
1947 replay_barrier $SINGLEMDS
1948 sleep 1 # give clients a time to do operations
1949 # Increment the number of failovers
1950 num_failovers=$((num_failovers+1))
1951 log "$TESTNAME fail $SINGLEMDS $num_failovers times"
1953 elapsed=$(($(date +%s) - start_ts))
1956 wait $pid || error "rundbench load on $clients failed!"
1958 run_test 70b "mds recovery; $CLIENTCOUNT clients"
1959 # end multi-client tests
1962 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1966 replay_barrier $SINGLEMDS
1967 #define OBD_FAIL_LDLM_ENQUEUE_NET 0x302
1968 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000302"
1971 wait $pid || return 1
1972 [ -e $DIR/$tfile ] && return 2
1975 run_test 73a "open(O_CREAT), unlink, replay, reconnect before open replay , close"
1978 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1982 replay_barrier $SINGLEMDS
1983 #define OBD_FAIL_LDLM_REPLY 0x30c
1984 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000030c"
1987 wait $pid || return 1
1988 [ -e $DIR/$tfile ] && return 2
1991 run_test 73b "open(O_CREAT), unlink, replay, reconnect at open_replay reply, close"
1994 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1998 replay_barrier $SINGLEMDS
1999 #define OBD_FAIL_TGT_LAST_REPLAY 0x710
2000 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000710"
2003 wait $pid || return 1
2004 [ -e $DIR/$tfile ] && return 2
2007 run_test 73c "open(O_CREAT), unlink, replay, reconnect at last_replay, close"
2011 local clients=${CLIENTS:-$HOSTNAME}
2013 zconf_umount_clients $clients $MOUNT
2015 facet_failover $SINGLEMDS
2016 zconf_mount_clients $clients $MOUNT
2018 touch $DIR/$tfile || return 1
2019 rm $DIR/$tfile || return 2
2020 clients_up || error "client evicted: $?"
2023 run_test 74 "Ensure applications don't fail waiting for OST recovery"
2025 remote_dir_check_80() {
2027 local diridx=$($GETSTRIPE -M $remote_dir)
2028 [ $diridx -eq $MDTIDX ] || error "$diridx != $MDTIDX"
2030 createmany -o $remote_dir/f-%d 20 || error "creation failed"
2031 local fileidx=$($GETSTRIPE -M $remote_dir/f-1)
2032 [ $fileidx -eq $MDTIDX ] || error "$fileidx != $MDTIDX"
2038 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2039 ([ $FAILURE_MODE == "HARD" ] &&
2040 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2041 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2045 local remote_dir=$DIR/$tdir/remote_dir
2048 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2049 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1701
2050 $LFS mkdir -i $MDTIDX $remote_dir &
2053 fail mds$((MDTIDX + 1))
2055 wait $CLIENT_PID || error "remote creation failed"
2057 remote_dir_check_80 || error "remote dir check failed"
2058 rm -rf $DIR/$tdir || error "rmdir failed"
2062 run_test 80a "DNE: create remote dir, drop update rep from MDT1, fail MDT1"
2065 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2066 ([ $FAILURE_MODE == "HARD" ] &&
2067 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2068 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2072 local remote_dir=$DIR/$tdir/remote_dir
2075 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2076 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1701
2077 $LFS mkdir -i $MDTIDX $remote_dir &
2082 wait $CLIENT_PID || error "remote creation failed"
2084 remote_dir_check_80 || error "remote dir check failed"
2085 rm -rf $DIR/$tdir || error "rmdir failed"
2089 run_test 80b "DNE: create remote dir, drop update rep from MDT1, fail MDT0"
2092 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2093 ([ $FAILURE_MODE == "HARD" ] &&
2094 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2095 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2099 local remote_dir=$DIR/$tdir/remote_dir
2102 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2103 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1701
2104 $LFS mkdir -i $MDTIDX $remote_dir &
2108 fail mds$((MDTIDX + 1))
2110 wait $CLIENT_PID || error "remote creation failed"
2112 remote_dir_check_80 || error "remote dir check failed"
2113 rm -rf $DIR/$tdir || error "rmdir failed"
2117 run_test 80c "DNE: create remote dir, drop update rep from MDT1, fail MDT[0,1]"
2120 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2122 local remote_dir=$DIR/$tdir/remote_dir
2125 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2126 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1701
2127 $LFS mkdir -i $MDTIDX $remote_dir &
2130 fail mds${MDTIDX},mds$((MDTIDX + 1))
2132 wait $CLIENT_PID || error "remote creation failed"
2134 remote_dir_check_80 || error "remote dir check failed"
2135 rm -rf $DIR/$tdir || error "rmdir failed"
2139 run_test 80d "DNE: create remote dir, drop update rep from MDT1, fail 2 MDTs"
2142 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2143 ([ $FAILURE_MODE == "HARD" ] &&
2144 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2145 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2149 local remote_dir=$DIR/$tdir/remote_dir
2152 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2153 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2154 $LFS mkdir -i $MDTIDX $remote_dir &
2159 wait $CLIENT_PID || error "remote creation failed"
2161 remote_dir_check_80 || error "remote dir check failed"
2162 rm -rf $DIR/$tdir || error "rmdir failed"
2166 run_test 80e "DNE: create remote dir, drop MDT0 rep, fail MDT0"
2169 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2170 ([ $FAILURE_MODE == "HARD" ] &&
2171 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2172 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2175 local remote_dir=$DIR/$tdir/remote_dir
2178 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2179 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2180 $LFS mkdir -i $MDTIDX $remote_dir &
2183 fail mds$((MDTIDX + 1))
2185 wait $CLIENT_PID || error "remote creation failed"
2187 remote_dir_check_80 || error "remote dir check failed"
2188 rm -rf $DIR/$tdir || error "rmdir failed"
2192 run_test 80f "DNE: create remote dir, drop MDT0 rep, fail MDT1"
2195 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2196 ([ $FAILURE_MODE == "HARD" ] &&
2197 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2198 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2202 local remote_dir=$DIR/$tdir/remote_dir
2205 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2206 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2207 $LFS mkdir -i $MDTIDX $remote_dir &
2211 fail mds$((MDTIDX + 1))
2213 wait $CLIENT_PID || error "remote creation failed"
2215 remote_dir_check_80 || error "remote dir check failed"
2216 rm -rf $DIR/$tdir || error "rmdir failed"
2220 run_test 80g "DNE: create remote dir, drop MDT0 rep, fail MDT0, then MDT1"
2223 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2225 local remote_dir=$DIR/$tdir/remote_dir
2228 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2229 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2230 $LFS mkdir -i $MDTIDX $remote_dir &
2233 fail mds${MDTIDX},mds$((MDTIDX + 1))
2235 wait $CLIENT_PID || return 1
2237 remote_dir_check_80 || error "remote dir check failed"
2238 rm -rf $DIR/$tdir || error "rmdir failed"
2242 run_test 80h "DNE: create remote dir, drop MDT0 rep, fail 2 MDTs"
2245 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2246 ([ $FAILURE_MODE == "HARD" ] &&
2247 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2248 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2252 local remote_dir=$DIR/$tdir/remote_dir
2255 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2258 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2259 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2263 fail mds$((MDTIDX + 1))
2265 wait $CLIENT_PID || error "rm remote dir failed"
2267 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2269 rm -rf $DIR/$tdir || error "rmdir failed"
2273 run_test 81a "DNE: unlink remote dir, drop MDT0 update rep, fail MDT1"
2276 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2277 ([ $FAILURE_MODE == "HARD" ] &&
2278 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2279 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2282 local remote_dir=$DIR/$tdir/remote_dir
2285 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2287 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2288 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2294 wait $CLIENT_PID || error "rm remote dir failed"
2296 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2298 rm -rf $DIR/$tdir || error "rmdir failed"
2302 run_test 81b "DNE: unlink remote dir, drop MDT0 update reply, fail MDT0"
2305 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2306 ([ $FAILURE_MODE == "HARD" ] &&
2307 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2308 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2312 local remote_dir=$DIR/$tdir/remote_dir
2315 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2317 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2318 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2323 fail mds$((MDTIDX + 1))
2325 wait $CLIENT_PID || error "rm remote dir failed"
2327 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2329 rm -rf $DIR/$tdir || error "rmdir failed"
2333 run_test 81c "DNE: unlink remote dir, drop MDT0 update reply, fail MDT0,MDT1"
2336 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2338 local remote_dir=$DIR/$tdir/remote_dir
2341 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2343 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2344 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2348 fail mds${MDTIDX},mds$((MDTIDX + 1))
2350 wait $CLIENT_PID || error "rm remote dir failed"
2352 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2354 rm -rf $DIR/$tdir || error "rmdir failed"
2358 run_test 81d "DNE: unlink remote dir, drop MDT0 update reply, fail 2 MDTs"
2361 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2362 ([ $FAILURE_MODE == "HARD" ] &&
2363 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2364 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2368 local remote_dir=$DIR/$tdir/remote_dir
2371 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2373 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2374 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2377 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0
2381 wait $CLIENT_PID || error "rm remote dir failed"
2383 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2385 rm -rf $DIR/$tdir || error "rmdir failed"
2389 run_test 81e "DNE: unlink remote dir, drop MDT1 req reply, fail MDT0"
2392 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2393 ([ $FAILURE_MODE == "HARD" ] &&
2394 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2395 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2399 local remote_dir=$DIR/$tdir/remote_dir
2402 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2404 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2405 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2409 fail mds$((MDTIDX + 1))
2411 wait $CLIENT_PID || error "rm remote dir failed"
2413 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2415 rm -rf $DIR/$tdir || error "rmdir failed"
2419 run_test 81f "DNE: unlink remote dir, drop MDT1 req reply, fail MDT1"
2422 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2423 ([ $FAILURE_MODE == "HARD" ] &&
2424 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2425 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2429 local remote_dir=$DIR/$tdir/remote_dir
2432 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2434 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2435 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2440 fail mds$((MDTIDX + 1))
2442 wait $CLIENT_PID || error "rm remote dir failed"
2444 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2446 rm -rf $DIR/$tdir || error "rmdir failed"
2450 run_test 81g "DNE: unlink remote dir, drop req reply, fail M0, then M1"
2453 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2455 local remote_dir=$DIR/$tdir/remote_dir
2458 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2460 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2461 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2465 fail mds${MDTIDX},mds$((MDTIDX + 1))
2467 wait $CLIENT_PID || error "rm remote dir failed"
2469 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2471 rm -rf $DIR/$tdir || error "rmdir failed"
2475 run_test 81h "DNE: unlink remote dir, drop request reply, fail 2 MDTs"
2479 createmany -o $DIR/$tdir/$tfile- 10 || return 1
2480 #define OBD_FAIL_MDS_FAIL_LOV_LOG_ADD 0x140
2481 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000140"
2482 unlinkmany $DIR/$tdir/$tfile- 10 || return 2
2484 run_test 83a "fail log_add during unlink recovery"
2488 createmany -o $DIR/$tdir/$tfile- 10 || return 1
2489 replay_barrier $SINGLEMDS
2490 unlinkmany $DIR/$tdir/$tfile- 10 || return 2
2491 #define OBD_FAIL_MDS_FAIL_LOV_LOG_ADD 0x140
2492 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000140"
2495 run_test 83b "fail log_add during unlink recovery"
2498 #define OBD_FAIL_MDS_OPEN_WAIT_CREATE 0x144
2499 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000144"
2500 createmany -o $DIR/$tfile- 1 &
2504 client_up || client_up || true # reconnect
2506 run_test 84a "stale open during export disconnect"
2508 test_85a() { #bug 16774
2509 lctl set_param -n ldlm.cancel_unused_locks_before_replay "1"
2511 for i in `seq 100`; do
2512 echo "tag-$i" > $DIR/$tfile-$i
2513 grep -q "tag-$i" $DIR/$tfile-$i || error "f2-$i"
2516 lov_id=`lctl dl | grep "clilov"`
2517 addr=`echo $lov_id | awk '{print $4}' | awk -F '-' '{print $3}'`
2518 count=`lctl get_param -n ldlm.namespaces.*MDT0000*$addr.lock_unused_count`
2519 echo "before recovery: unused locks count = $count"
2523 count2=`lctl get_param -n ldlm.namespaces.*MDT0000*$addr.lock_unused_count`
2524 echo "after recovery: unused locks count = $count2"
2526 if [ $count2 -ge $count ]; then
2527 error "unused locks are not canceled"
2530 run_test 85a "check the cancellation of unused locks during recovery(IBITS)"
2532 test_85b() { #bug 16774
2533 lctl set_param -n ldlm.cancel_unused_locks_before_replay "1"
2535 do_facet mgs $LCTL pool_new $FSNAME.$TESTNAME || return 1
2536 do_facet mgs $LCTL pool_add $FSNAME.$TESTNAME $FSNAME-OST0000 || return 2
2538 $SETSTRIPE -c 1 -p $FSNAME.$TESTNAME $DIR
2540 for i in `seq 100`; do
2541 dd if=/dev/urandom of=$DIR/$tfile-$i bs=4096 count=32 >/dev/null 2>&1
2544 cancel_lru_locks osc
2546 for i in `seq 100`; do
2547 dd if=$DIR/$tfile-$i of=/dev/null bs=4096 count=32 >/dev/null 2>&1
2550 lov_id=`lctl dl | grep "clilov"`
2551 addr=`echo $lov_id | awk '{print $4}' | awk -F '-' '{print $3}'`
2552 count=`lctl get_param -n ldlm.namespaces.*OST0000*$addr.lock_unused_count`
2553 echo "before recovery: unused locks count = $count"
2554 [ $count != 0 ] || return 3
2558 count2=`lctl get_param -n ldlm.namespaces.*OST0000*$addr.lock_unused_count`
2559 echo "after recovery: unused locks count = $count2"
2561 do_facet mgs $LCTL pool_remove $FSNAME.$TESTNAME $FSNAME-OST0000 || return 4
2562 do_facet mgs $LCTL pool_destroy $FSNAME.$TESTNAME || return 5
2564 if [ $count2 -ge $count ]; then
2565 error "unused locks are not canceled"
2568 run_test 85b "check the cancellation of unused locks during recovery(EXTENT)"
2571 local clients=${CLIENTS:-$HOSTNAME}
2573 zconf_umount_clients $clients $MOUNT
2574 do_facet $SINGLEMDS lctl set_param mdt.${FSNAME}-MDT*.exports.clear=0
2575 remount_facet $SINGLEMDS
2576 zconf_mount_clients $clients $MOUNT
2578 run_test 86 "umount server after clear nid_stats should not hit LBUG"
2581 do_facet ost1 "lctl set_param -n obdfilter.${ost1_svc}.sync_journal 0"
2584 $SETSTRIPE -i 0 -c 1 $DIR/$tfile
2585 dd if=/dev/urandom of=$DIR/$tfile bs=1024k count=8 || error "Cannot write"
2586 cksum=`md5sum $DIR/$tfile | awk '{print $1}'`
2587 cancel_lru_locks osc
2589 dd if=$DIR/$tfile of=/dev/null bs=1024k count=8 || error "Cannot read"
2590 cksum2=`md5sum $DIR/$tfile | awk '{print $1}'`
2591 if [ $cksum != $cksum2 ] ; then
2592 error "New checksum $cksum2 does not match original $cksum"
2595 run_test 87 "write replay"
2598 do_facet ost1 "lctl set_param -n obdfilter.${ost1_svc}.sync_journal 0"
2601 $SETSTRIPE -i 0 -c 1 $DIR/$tfile
2602 dd if=/dev/urandom of=$DIR/$tfile bs=1024k count=8 || error "Cannot write"
2603 sleep 1 # Give it a chance to flush dirty data
2604 echo TESTTEST | dd of=$DIR/$tfile bs=1 count=8 seek=64
2605 cksum=`md5sum $DIR/$tfile | awk '{print $1}'`
2606 cancel_lru_locks osc
2608 dd if=$DIR/$tfile of=/dev/null bs=1024k count=8 || error "Cannot read"
2609 cksum2=`md5sum $DIR/$tfile | awk '{print $1}'`
2610 if [ $cksum != $cksum2 ] ; then
2611 error "New checksum $cksum2 does not match original $cksum"
2614 run_test 87b "write replay with changed data (checksum resend)"
2616 test_88() { #bug 17485
2620 $SETSTRIPE -i 0 -c 1 $DIR/$tdir || error "$SETSTRIPE"
2623 replay_barrier $SINGLEMDS
2625 # exhaust precreations on ost1
2626 local OST=$(ostname_from_index 0)
2627 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $OST)
2628 local last_id=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_last_id)
2629 local next_id=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_next_id)
2630 echo "before test: last_id = $last_id, next_id = $next_id"
2632 echo "Creating to objid $last_id on ost $OST..."
2633 createmany -o $DIR/$tdir/f-%d $next_id $((last_id - next_id + 2))
2635 #create some files to use some uncommitted objids
2636 last_id=$(($last_id + 1))
2637 createmany -o $DIR/$tdir/f-%d $last_id 8
2639 last_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_last_id)
2640 next_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_next_id)
2641 echo "before recovery: last_id = $last_id2, next_id = $next_id2"
2643 # if test uses shutdown_facet && reboot_facet instead of facet_failover ()
2644 # it has to take care about the affected facets, bug20407
2645 local affected_mds1=$(affected_facets mds1)
2646 local affected_ost1=$(affected_facets ost1)
2648 shutdown_facet $SINGLEMDS
2651 reboot_facet $SINGLEMDS
2652 change_active $affected_mds1
2653 wait_for_facet $affected_mds1
2654 mount_facets $affected_mds1 || error "Restart of mds failed"
2657 change_active $affected_ost1
2658 wait_for_facet $affected_ost1
2659 mount_facets $affected_ost1 || error "Restart of ost1 failed"
2663 last_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_last_id)
2664 next_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_next_id)
2665 echo "after recovery: last_id = $last_id2, next_id = $next_id2"
2667 # create new files, which should use new objids, and ensure the orphan
2668 # cleanup phase for ost1 is completed at the same time
2669 for i in `seq 8`; do
2670 file_id=$(($last_id + 10 + $i))
2671 dd if=/dev/urandom of=$DIR/$tdir/f-$file_id bs=4096 count=128
2674 # if the objids were not recreated, then "ls" will failed for -ENOENT
2675 ls -l $DIR/$tdir/* || error "can't get the status of precreated files"
2678 # write into previously created files
2679 for i in `seq 8`; do
2680 file_id=$(($last_id + $i))
2681 dd if=/dev/urandom of=$DIR/$tdir/f-$file_id bs=4096 count=128
2682 cp -f $DIR/$tdir/f-$file_id $TMP/$tdir/
2685 # compare the content
2686 for i in `seq 8`; do
2687 file_id=$(($last_id + $i))
2688 cmp $TMP/$tdir/f-$file_id $DIR/$tdir/f-$file_id || error "the content" \
2689 "of file is modified!"
2694 run_test 88 "MDS should not assign same objid to different files "
2697 cancel_lru_locks osc
2699 rm -f $DIR/$tdir/$tfile
2701 wait_delete_completed
2702 BLOCKS1=$(df -P $MOUNT | tail -n 1 | awk '{ print $3 }')
2703 $SETSTRIPE -i 0 -c 1 $DIR/$tdir/$tfile
2704 dd if=/dev/zero bs=1M count=10 of=$DIR/$tdir/$tfile
2707 facet_failover $SINGLEMDS
2708 rm $DIR/$tdir/$tfile
2711 zconf_mount $(hostname) $MOUNT
2712 client_up || return 1
2714 wait_delete_completed
2715 BLOCKS2=$(df -P $MOUNT | tail -n 1 | awk '{ print $3 }')
2716 [ $((BLOCKS2 - BLOCKS1)) -le 4 ] || \
2717 error $((BLOCKS2 - BLOCKS1)) blocks leaked
2720 run_test 89 "no disk space leak on late ost connection"
2726 change_active $facet
2727 wait_for_facet $facet
2728 mount_facet $facet || error "Restart of $facet failed"
2732 test_90() { # bug 19494
2733 local dir=$DIR/$tdir
2734 local ostfail=$(get_random_entry $(get_facets OST))
2736 if [[ $FAILURE_MODE = HARD ]]; then
2737 local affected=$(affected_facets $ostfail);
2738 if [[ "$affected" != $ostfail ]]; then
2739 skip not functional with FAILURE_MODE=$FAILURE_MODE, affected: $affected
2746 echo "Create the files"
2748 # file "f${index}" striped over 1 OST
2749 # file "all" striped over all OSTs
2751 $SETSTRIPE -c $OSTCOUNT $dir/all ||
2752 error "setstripe failed to create $dir/all"
2754 for (( i=0; i<$OSTCOUNT; i++ )); do
2756 $SETSTRIPE -i $i -c 1 $f || error "$SETSTRIPE failed to create $f"
2758 # confirm setstripe actually created the stripe on the requested OST
2759 local uuid=$(ostuuid_from_index $i)
2760 for file in f$i all; do
2761 if [[ $dir/$file != $($LFS find --obd $uuid --name $file $dir) ]]; then
2762 $GETSTRIPE $dir/$file
2763 error wrong stripe: $file, uuid: $uuid
2768 # Before failing an OST, get its obd name and index
2769 local varsvc=${ostfail}_svc
2770 local obd=$(do_facet $ostfail lctl get_param -n obdfilter.${!varsvc}.uuid)
2771 local index=$(($(facet_number $ostfail) - 1))
2773 echo "Fail $ostfail $obd, display the list of affected files"
2774 shutdown_facet $ostfail || return 2
2776 trap "cleanup_90 $ostfail" EXIT INT
2777 echo "General Query: lfs find $dir"
2778 local list=$($LFS find $dir)
2780 for (( i=0; i<$OSTCOUNT; i++ )); do
2781 list_member "$list" $dir/f$i || error_noexit "lfs find $dir: no file f$i"
2783 list_member "$list" $dir/all || error_noexit "lfs find $dir: no file all"
2785 # focus on the missing OST,
2786 # we expect to see only two files affected: "f$(index)" and "all"
2788 echo "Querying files on shutdown $ostfail: lfs find --obd $obd"
2789 list=$($LFS find --obd $obd $dir)
2791 for file in all f$index; do
2792 list_member "$list" $dir/$file ||
2793 error_noexit "lfs find does not report the affected $obd for $file"
2796 [[ $(echo $list | wc -w) -eq 2 ]] ||
2797 error_noexit "lfs find reports the wrong list of affected files ${#list[@]}"
2799 echo "Check getstripe: $GETSTRIPE -r --obd $obd"
2800 list=$($GETSTRIPE -r --obd $obd $dir)
2802 for file in all f$index; do
2803 echo "$list" | grep $dir/$file ||
2804 error_noexit "lfs getsripe does not report the affected $obd for $file"
2809 run_test 90 "lfs find identifies the missing striped file segments"
2812 local server_version=$(lustre_version_code $SINGLEMDS)
2813 [[ $server_version -ge $(version_code 2.6.90) ]] ||
2814 [[ $server_version -ge $(version_code 2.5.4) &&
2815 $server_version -lt $(version_code 2.5.50) ]] ||
2816 { skip "Need MDS version 2.5.4+ or 2.6.90+"; return; }
2818 cancel_lru_locks osc
2820 $SETSTRIPE -i 0 -c 1 $DIR/$tfile
2821 dd if=/dev/zero of=$DIR/$tfile bs=1024 count=1
2822 #define OBD_FAIL_TGT_REPLAY_RECONNECT 0x715
2823 # We need to emulate a state that OST is waiting for other clients
2824 # not completing the recovery. Final ping is queued, but reply will be sent
2825 # on the recovery completion. It is done by sleep before processing final
2827 do_facet ost1 "$LCTL set_param fail_val=40"
2828 do_facet ost1 "$LCTL set_param fail_loc=0x715"
2831 run_test 93 "replay + reconnect"
2834 check_and_cleanup_lustre