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 BEFOREUSED=`df -P $DIR | tail -1 | awk '{ print $3 }'`
429 dd if=/dev/zero of=$DIR/$tfile bs=4k count=10000 &
431 while [ ! -e $DIR/$tfile ] ; do
432 usleep 60 # give dd a chance to start
435 $GETSTRIPE $DIR/$tfile || return 1
436 rm -f $DIR/$tfile || return 2 # make it an orphan
438 client_up || client_up || true # reconnect
440 fail $SINGLEMDS # start orphan recovery
441 wait_recovery_complete $SINGLEMDS || error "MDS recovery not done"
442 wait_mds_ost_sync || return 3
443 AFTERUSED=`df -P $DIR | tail -1 | awk '{ print $3 }'`
444 log "before $BEFOREUSED, after $AFTERUSED"
445 (( $AFTERUSED > $BEFOREUSED + $(fs_log_size) )) &&
446 error "after $AFTERUSED > before $BEFOREUSED"
449 run_test 20b "write, unlink, eviction, replay, (test mds_cleanup_orphans)"
451 test_20c() { # bug 10480
452 multiop_bg_pause $DIR/$tfile Ow_c || return 1
458 client_up || client_up || true # reconnect
461 wait $pid || return 1
462 [ -s $DIR/$tfile ] || error "File was truncated"
466 run_test 20c "check that client eviction does not affect file content"
469 replay_barrier $SINGLEMDS
470 multiop_bg_pause $DIR/$tfile O_tSc || return 5
473 touch $DIR/g11 || return 1
477 wait $pid || return 2
478 [ -e $DIR/$tfile ] && return 3
479 touch $DIR/h11 || return 4
482 run_test 21 "|X| open(O_CREAT), unlink touch new, replay, close (test mds_cleanup_orphans)"
485 multiop_bg_pause $DIR/$tfile O_tSc || return 3
488 replay_barrier $SINGLEMDS
493 wait $pid || return 1
494 [ -e $DIR/$tfile ] && return 2
497 run_test 22 "open(O_CREAT), |X| unlink, replay, close (test mds_cleanup_orphans)"
500 multiop_bg_pause $DIR/$tfile O_tSc || return 5
503 replay_barrier $SINGLEMDS
505 touch $DIR/g11 || return 1
509 wait $pid || return 2
510 [ -e $DIR/$tfile ] && return 3
511 touch $DIR/h11 || return 4
514 run_test 23 "open(O_CREAT), |X| unlink touch new, replay, close (test mds_cleanup_orphans)"
517 multiop_bg_pause $DIR/$tfile O_tSc || return 3
520 replay_barrier $SINGLEMDS
524 wait $pid || return 1
525 [ -e $DIR/$tfile ] && return 2
528 run_test 24 "open(O_CREAT), replay, unlink, close (test mds_cleanup_orphans)"
531 multiop_bg_pause $DIR/$tfile O_tSc || return 3
535 replay_barrier $SINGLEMDS
538 wait $pid || return 1
539 [ -e $DIR/$tfile ] && return 2
542 run_test 25 "open(O_CREAT), unlink, replay, close (test mds_cleanup_orphans)"
545 replay_barrier $SINGLEMDS
546 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
548 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
553 wait $pid2 || return 1
557 wait $pid1 || return 2
558 [ -e $DIR/$tfile-1 ] && return 3
559 [ -e $DIR/$tfile-2 ] && return 4
562 run_test 26 "|X| open(O_CREAT), unlink two, close one, replay, close one (test mds_cleanup_orphans)"
565 replay_barrier $SINGLEMDS
566 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
568 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
575 wait $pid1 || return 1
577 wait $pid2 || return 2
578 [ -e $DIR/$tfile-1 ] && return 3
579 [ -e $DIR/$tfile-2 ] && return 4
582 run_test 27 "|X| open(O_CREAT), unlink two, replay, close two (test mds_cleanup_orphans)"
585 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
587 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
589 replay_barrier $SINGLEMDS
593 wait $pid2 || return 1
597 wait $pid1 || return 2
598 [ -e $DIR/$tfile-1 ] && return 3
599 [ -e $DIR/$tfile-2 ] && return 4
602 run_test 28 "open(O_CREAT), |X| unlink two, close one, replay, close one (test mds_cleanup_orphans)"
605 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
607 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
609 replay_barrier $SINGLEMDS
615 wait $pid1 || return 1
617 wait $pid2 || return 2
618 [ -e $DIR/$tfile-1 ] && return 3
619 [ -e $DIR/$tfile-2 ] && return 4
622 run_test 29 "open(O_CREAT), |X| unlink two, replay, close two (test mds_cleanup_orphans)"
625 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
627 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
632 replay_barrier $SINGLEMDS
635 wait $pid1 || return 1
637 wait $pid2 || return 2
638 [ -e $DIR/$tfile-1 ] && return 3
639 [ -e $DIR/$tfile-2 ] && return 4
642 run_test 30 "open(O_CREAT) two, unlink two, replay, close two (test mds_cleanup_orphans)"
645 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
647 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
651 replay_barrier $SINGLEMDS
655 wait $pid1 || return 1
657 wait $pid2 || return 2
658 [ -e $DIR/$tfile-1 ] && return 3
659 [ -e $DIR/$tfile-2 ] && return 4
662 run_test 31 "open(O_CREAT) two, unlink one, |X| unlink one, close two (test mds_cleanup_orphans)"
664 # tests for bug 2104; completion without crashing is success. The close is
665 # stale, but we always return 0 for close, so the app never sees it.
667 multiop_bg_pause $DIR/$tfile O_c || return 2
669 multiop_bg_pause $DIR/$tfile O_c || return 3
672 client_up || client_up || return 1
675 wait $pid1 || return 4
676 wait $pid2 || return 5
679 run_test 32 "close() notices client eviction; close() after client eviction"
682 createmany -o $DIR/$tfile-%d 10
683 replay_barrier_nosync $SINGLEMDS
684 fail_abort $SINGLEMDS
685 # recreate shouldn't fail
686 createmany -o $DIR/$tfile--%d 10 || return 1
690 run_test 33a "fid seq shouldn't be reused after abort recovery"
693 #define OBD_FAIL_SEQ_ALLOC 0x1311
694 do_facet $SINGLEMDS "lctl set_param fail_loc=0x1311"
696 createmany -o $DIR/$tfile-%d 10
697 replay_barrier_nosync $SINGLEMDS
698 fail_abort $SINGLEMDS
699 # recreate shouldn't fail
700 createmany -o $DIR/$tfile--%d 10 || return 1
704 run_test 33b "test fid seq allocation"
707 multiop_bg_pause $DIR/$tfile O_c || return 2
711 replay_barrier $SINGLEMDS
712 fail_abort $SINGLEMDS
714 wait $pid || return 3
715 [ -e $DIR/$tfile ] && return 1
719 run_test 34 "abort recovery before client does replay (test mds_cleanup_orphans)"
721 # bug 2278 - generate one orphan on OST, then destroy it during recovery from llog
725 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
726 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000119"
731 # give a chance to remove from MDS
732 fail_abort $SINGLEMDS
733 $CHECKSTAT -t file $DIR/$tfile && return 1 || true
735 run_test 35 "test recovery from llog for unlink op"
737 # b=2432 resent cancel after replay uses wrong cookie,
738 # so don't resend cancels
740 replay_barrier $SINGLEMDS
742 checkstat $DIR/$tfile
743 facet_failover $SINGLEMDS
745 if dmesg | grep "unknown lock cookie"; then
746 echo "cancel after replay failed"
750 run_test 36 "don't resend cancel"
753 # directory orphans can't be unlinked from PENDING directory
755 rmdir $DIR/$tfile 2>/dev/null
756 multiop_bg_pause $DIR/$tfile dD_c || return 2
760 replay_barrier $SINGLEMDS
761 # clear the dmesg buffer so we only see errors from this recovery
762 do_facet $SINGLEMDS dmesg -c >/dev/null
763 fail_abort $SINGLEMDS
765 do_facet $SINGLEMDS dmesg | grep "error .* unlinking .* from PENDING" &&
767 wait $pid || return 3
771 start_full_debug_logging
772 run_test 37 "abort recovery before client does replay (test mds_cleanup_orphans for directories)"
773 stop_full_debug_logging
776 createmany -o $DIR/$tfile-%d 800
777 unlinkmany $DIR/$tfile-%d 0 400
778 replay_barrier $SINGLEMDS
780 unlinkmany $DIR/$tfile-%d 400 400
782 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
784 run_test 38 "test recovery from unlink llog (test llog_gen_rec) "
786 test_39() { # bug 4176
787 createmany -o $DIR/$tfile-%d 800
788 replay_barrier $SINGLEMDS
789 unlinkmany $DIR/$tfile-%d 0 400
791 unlinkmany $DIR/$tfile-%d 400 400
793 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
795 run_test 39 "test recovery from unlink llog (test llog_gen_rec) "
798 lctl get_param -n osc.*.stats | awk -vwrites=0 '/ost_write/ { writes += $2 } END { print writes; }'
803 # always need connection to MDS to verify layout during IO. LU-2628.
804 lctl get_param mdc.*.connect_flags | grep -q layout_lock &&
805 skip "layout_lock needs MDS connection for IO" && return 0
807 $LCTL mark multiop $MOUNT/$tfile OS_c
808 multiop $MOUNT/$tfile OS_c &
810 writeme -s $MOUNT/${tfile}-2 &
813 facet_failover $SINGLEMDS
814 #define OBD_FAIL_MDS_CONNECT_NET 0x117
815 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000117"
817 stat1=`count_ost_writes`
819 stat2=`count_ost_writes`
820 echo "$stat1, $stat2"
821 if [ $stat1 -lt $stat2 ]; then
822 echo "writes continuing during recovery"
825 echo "writes not continuing during recovery, bug 2477"
828 echo "waiting for writeme $WRITE_PID"
832 echo "waiting for multiop $PID"
833 wait $PID || return 2
834 do_facet client munlink $MOUNT/$tfile || return 3
835 do_facet client munlink $MOUNT/${tfile}-2 || return 3
838 run_test 40 "cause recovery in ptlrpc, ensure IO continues"
842 # make sure that a read to one osc doesn't try to double-unlock its page just
843 # because another osc is invalid. trigger_group_io used to mistakenly return
844 # an error if any oscs were invalid even after having successfully put rpcs
845 # on valid oscs. This was fatal if the caller was ll_readpage who unlocked
846 # the page, guarnateeing that the unlock from the RPC completion would
847 # assert on trying to unlock the unlocked page.
849 [ $OSTCOUNT -lt 2 ] &&
850 skip_env "skipping test 41: we don't have a second OST to test with" &&
853 local f=$MOUNT/$tfile
854 # make sure the start of the file is ost1
855 $SETSTRIPE -S $((128 * 1024)) -i 0 $f
856 do_facet client dd if=/dev/zero of=$f bs=4k count=1 || return 3
858 # fail ost2 and read from ost1
859 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $ost2_svc)
860 local osc2dev=$(do_facet $SINGLEMDS "lctl get_param -n devices" | \
861 grep $mdtosc | awk '{print $1}')
862 [ -z "$osc2dev" ] && echo "OST: $ost2_svc" && lctl get_param -n devices &&
864 do_facet $SINGLEMDS $LCTL --device $osc2dev deactivate || return 1
865 do_facet client dd if=$f of=/dev/null bs=4k count=1 || return 3
866 do_facet $SINGLEMDS $LCTL --device $osc2dev activate || return 2
869 run_test 41 "read from a valid osc while other oscs are invalid"
871 # test MDS recovery after ost failure
873 blocks=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
874 createmany -o $DIR/$tfile-%d 800
876 unlinkmany $DIR/$tfile-%d 0 400
878 lctl set_param debug=-1
881 # osc is evicted, fs is smaller (but only with failout OSTs (bug 7287)
882 #blocks_after=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
883 #[ $blocks_after -lt $blocks ] || return 1
884 echo wait for MDS to timeout and recover
885 sleep $((TIMEOUT * 2))
887 unlinkmany $DIR/$tfile-%d 400 400
888 $CHECKSTAT -t file $DIR/$tfile-* && return 2 || true
890 run_test 42 "recovery after ost failure"
892 # timeout in MDS/OST recovery RPC will LBUG MDS
893 test_43() { # bug 2530
894 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
896 replay_barrier $SINGLEMDS
898 # OBD_FAIL_OST_CREATE_NET 0x204
899 do_facet ost1 "lctl set_param fail_loc=0x80000204"
902 do_facet ost1 "lctl set_param fail_loc=0"
906 run_test 43 "mds osc import failure during recovery; don't LBUG"
908 test_44a() { # was test_44
911 local mdcdev=$($LCTL get_param -n devices |
912 awk "/ ${FSNAME}-MDT0000-mdc-/ {print \$1}")
913 [ "$mdcdev" ] || return 2
914 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
915 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
917 # adaptive timeouts slow this way down
918 if at_is_enabled; then
919 at_max_saved=$(at_max_get mds)
923 for i in `seq 1 10`; do
924 echo "$i of 10 ($(date +%s))"
925 do_facet $SINGLEMDS "lctl get_param -n mdt.*.mdt.timeouts | grep service"
926 #define OBD_FAIL_TGT_CONN_RACE 0x701
927 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000701"
928 # lctl below may fail, it is valid case
929 $LCTL --device $mdcdev recover
932 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
933 [ $at_max_saved -ne 0 ] && at_max_set $at_max_saved mds
936 run_test 44a "race in target handle connect"
939 local mdcdev=$($LCTL get_param -n devices |
940 awk "/ ${FSNAME}-MDT0000-mdc-/ {print \$1}")
941 [ "$mdcdev" ] || return 2
942 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
943 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
945 for i in `seq 1 10`; do
946 echo "$i of 10 ($(date +%s))"
947 do_facet $SINGLEMDS "lctl get_param -n mdt.*.mdt.timeouts | grep service"
948 #define OBD_FAIL_TGT_DELAY_RECONNECT 0x704
949 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000704"
950 # lctl below may fail, it is valid case
951 $LCTL --device $mdcdev recover
954 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
957 run_test 44b "race in target handle connect"
960 replay_barrier $SINGLEMDS
961 createmany -m $DIR/$tfile-%d 100 || error "failed to create directories"
962 #define OBD_FAIL_TGT_RCVG_FLAG 0x712
963 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000712"
964 fail_abort $SINGLEMDS
965 unlinkmany $DIR/$tfile-%d 100 && error "unliked after fail abort"
967 unlinkmany $DIR/$tfile-%d 100 && error "unliked after fail"
970 run_test 44c "race in target handle connect"
972 # Handle failed close
974 local mdcdev=$($LCTL get_param -n devices |
975 awk "/ ${FSNAME}-MDT0000-mdc-/ {print \$1}")
976 [ "$mdcdev" ] || return 2
977 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
978 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
980 $LCTL --device $mdcdev recover || return 6
982 multiop_bg_pause $DIR/$tfile O_c || return 1
985 # This will cause the CLOSE to fail before even
986 # allocating a reply buffer
987 $LCTL --device $mdcdev deactivate || return 4
991 wait $pid || return 1
993 $LCTL --device $mdcdev activate || return 5
996 $CHECKSTAT -t file $DIR/$tfile || return 2
999 run_test 45 "Handle failed close"
1003 drop_reply "touch $DIR/$tfile"
1005 # ironically, the previous test, 45, will cause a real forced close,
1006 # so just look for one for this test
1007 dmesg | grep -i "force closing client file handle for $tfile" && return 1
1010 run_test 46 "Don't leak file handle after open resend (3325)"
1012 test_47() { # bug 2824
1013 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1015 # create some files to make sure precreate has been done on all
1016 # OSTs. (just in case this test is run independently)
1017 createmany -o $DIR/$tfile 20 || return 1
1019 # OBD_FAIL_OST_CREATE_NET 0x204
1021 do_facet ost1 "lctl set_param fail_loc=0x80000204"
1022 client_up || return 2
1024 # let the MDS discover the OST failure, attempt to recover, fail
1025 # and recover again.
1026 sleep $((3 * TIMEOUT))
1028 # Without 2824, this createmany would hang
1029 createmany -o $DIR/$tfile 20 || return 3
1030 unlinkmany $DIR/$tfile 20 || return 4
1032 do_facet ost1 "lctl set_param fail_loc=0"
1035 run_test 47 "MDS->OSC failure during precreate cleanup (2824)"
1038 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1039 [ "$OSTCOUNT" -lt "2" ] && skip_env "$OSTCOUNT < 2 OSTs -- skipping" && return
1041 replay_barrier $SINGLEMDS
1042 createmany -o $DIR/$tfile 20 || return 1
1043 # OBD_FAIL_OST_EROFS 0x216
1044 facet_failover $SINGLEMDS
1045 do_facet ost1 "lctl set_param fail_loc=0x80000216"
1046 client_up || return 2
1048 createmany -o $DIR/$tfile 20 20 || return 2
1049 unlinkmany $DIR/$tfile 40 || return 3
1052 run_test 48 "MDS->OSC failure during precreate cleanup (2824)"
1055 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $ost1_svc)
1056 local oscdev=$(do_facet $SINGLEMDS "lctl get_param -n devices" | \
1057 grep $mdtosc | awk '{print $1}')
1058 [ "$oscdev" ] || return 1
1059 do_facet $SINGLEMDS $LCTL --device $oscdev recover || return 2
1060 do_facet $SINGLEMDS $LCTL --device $oscdev recover || return 3
1061 # give the mds_lov_sync threads a chance to run
1064 run_test 50 "Double OSC recovery, don't LASSERT (3812)"
1066 # b3764 timed out lock replay
1069 cancel_lru_locks mdc
1071 multiop $DIR/$tfile s || return 1
1072 replay_barrier $SINGLEMDS
1073 #define OBD_FAIL_LDLM_REPLY 0x30c
1074 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000030c"
1075 fail $SINGLEMDS || return 2
1076 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1078 $CHECKSTAT -t file $DIR/$tfile-* && return 3 || true
1080 run_test 52 "time out lock replay (3764)"
1082 # bug 3462 - simultaneous MDC requests
1084 cancel_lru_locks mdc # cleanup locks from former test cases
1085 mkdir -p $DIR/${tdir}-1
1086 mkdir -p $DIR/${tdir}-2
1087 multiop $DIR/${tdir}-1/f O_c &
1089 # give multiop a change to open
1092 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1093 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1094 kill -USR1 $close_pid
1095 cancel_lru_locks mdc # force the close
1096 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1098 mcreate $DIR/${tdir}-2/f || return 1
1100 # close should still be here
1101 [ -d /proc/$close_pid ] || return 2
1103 replay_barrier_nodf $SINGLEMDS
1105 wait $close_pid || return 3
1107 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1108 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1109 rm -rf $DIR/${tdir}-*
1111 run_test 53a "|X| close request while two MDC requests in flight"
1114 cancel_lru_locks mdc # cleanup locks from former test cases
1115 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1117 mkdir -p $DIR/${tdir}-1
1118 mkdir -p $DIR/${tdir}-2
1119 multiop_bg_pause $DIR/${tdir}-1/f O_c || return 6
1122 #define OBD_FAIL_MDS_REINT_NET 0x107
1123 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1124 mcreate $DIR/${tdir}-2/f &
1128 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1129 kill -USR1 $close_pid
1130 cancel_lru_locks mdc # force the close
1131 wait $close_pid || return 1
1132 # open should still be here
1133 [ -d /proc/$open_pid ] || return 2
1135 replay_barrier_nodf $SINGLEMDS
1137 wait $open_pid || return 3
1139 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1140 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1141 rm -rf $DIR/${tdir}-*
1143 run_test 53b "|X| open request while two MDC requests in flight"
1146 cancel_lru_locks mdc # cleanup locks from former test cases
1147 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1149 mkdir -p $DIR/${tdir}-1
1150 mkdir -p $DIR/${tdir}-2
1151 multiop $DIR/${tdir}-1/f O_c &
1154 #define OBD_FAIL_MDS_REINT_NET 0x107
1155 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1156 mcreate $DIR/${tdir}-2/f &
1160 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1161 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1162 kill -USR1 $close_pid
1163 cancel_lru_locks mdc # force the close
1165 #bz20647: make sure all pids are exists before failover
1166 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1167 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1168 replay_barrier_nodf $SINGLEMDS
1169 fail_nodf $SINGLEMDS
1170 wait $open_pid || return 1
1172 # close should be gone
1173 [ -d /proc/$close_pid ] && return 2
1174 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1176 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1177 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1178 rm -rf $DIR/${tdir}-*
1180 run_test 53c "|X| open request and close request while two MDC requests in flight"
1183 cancel_lru_locks mdc # cleanup locks from former test cases
1184 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1186 mkdir -p $DIR/${tdir}-1
1187 mkdir -p $DIR/${tdir}-2
1188 multiop $DIR/${tdir}-1/f O_c &
1190 # give multiop a chance to open
1193 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1194 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1195 kill -USR1 $close_pid
1196 cancel_lru_locks mdc # force the close
1197 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1198 mcreate $DIR/${tdir}-2/f || return 1
1200 # close should still be here
1201 [ -d /proc/$close_pid ] || return 2
1203 wait $close_pid || return 3
1205 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1206 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1207 rm -rf $DIR/${tdir}-*
1209 run_test 53d "|X| close reply while two MDC requests in flight"
1212 cancel_lru_locks mdc # cleanup locks from former test cases
1213 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1215 mkdir -p $DIR/${tdir}-1
1216 mkdir -p $DIR/${tdir}-2
1217 multiop $DIR/${tdir}-1/f O_c &
1220 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1221 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1222 mcreate $DIR/${tdir}-2/f &
1226 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1227 kill -USR1 $close_pid
1228 cancel_lru_locks mdc # force the close
1229 wait $close_pid || return 1
1230 # open should still be here
1231 [ -d /proc/$open_pid ] || return 2
1233 replay_barrier_nodf $SINGLEMDS
1235 wait $open_pid || return 3
1237 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1238 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1239 rm -rf $DIR/${tdir}-*
1241 run_test 53e "|X| open reply while two MDC requests in flight"
1244 cancel_lru_locks mdc # cleanup locks from former test cases
1245 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1247 mkdir -p $DIR/${tdir}-1
1248 mkdir -p $DIR/${tdir}-2
1249 multiop $DIR/${tdir}-1/f O_c &
1252 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1253 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1254 mcreate $DIR/${tdir}-2/f &
1258 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1259 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1260 kill -USR1 $close_pid
1261 cancel_lru_locks mdc # force the close
1263 #bz20647: make sure all pids are exists before failover
1264 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1265 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1266 replay_barrier_nodf $SINGLEMDS
1267 fail_nodf $SINGLEMDS
1268 wait $open_pid || return 1
1270 # close should be gone
1271 [ -d /proc/$close_pid ] && return 2
1272 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1274 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1275 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1276 rm -rf $DIR/${tdir}-*
1278 run_test 53f "|X| open reply and close reply while two MDC requests in flight"
1281 cancel_lru_locks mdc # cleanup locks from former test cases
1282 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1284 mkdir -p $DIR/${tdir}-1
1285 mkdir -p $DIR/${tdir}-2
1286 multiop $DIR/${tdir}-1/f O_c &
1289 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1290 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1291 mcreate $DIR/${tdir}-2/f &
1295 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1296 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1297 kill -USR1 $close_pid
1298 cancel_lru_locks mdc # force the close
1299 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1301 #bz20647: make sure all pids are exists before failover
1302 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1303 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1304 replay_barrier_nodf $SINGLEMDS
1305 fail_nodf $SINGLEMDS
1306 wait $open_pid || return 1
1308 # close should be gone
1309 [ -d /proc/$close_pid ] && return 2
1311 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1312 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1313 rm -rf $DIR/${tdir}-*
1315 run_test 53g "|X| drop open reply and close request while close and open are both in flight"
1318 cancel_lru_locks mdc # cleanup locks from former test cases
1319 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1321 mkdir -p $DIR/${tdir}-1
1322 mkdir -p $DIR/${tdir}-2
1323 multiop $DIR/${tdir}-1/f O_c &
1326 #define OBD_FAIL_MDS_REINT_NET 0x107
1327 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1328 mcreate $DIR/${tdir}-2/f &
1332 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1333 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1334 kill -USR1 $close_pid
1335 cancel_lru_locks mdc # force the close
1338 #bz20647: make sure all pids are exists before failover
1339 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1340 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1341 replay_barrier_nodf $SINGLEMDS
1342 fail_nodf $SINGLEMDS
1343 wait $open_pid || return 1
1345 # close should be gone
1346 [ -d /proc/$close_pid ] && return 2
1347 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1349 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1350 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1351 rm -rf $DIR/${tdir}-*
1353 run_test 53h "|X| open request and close reply while two MDC requests in flight"
1355 #b_cray 54 "|X| open request and close reply while two MDC requests in flight"
1357 #b3761 ASSERTION(hash != 0) failed
1359 # OBD_FAIL_MDS_OPEN_CREATE | OBD_FAIL_ONCE
1360 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012b"
1362 # give touch a chance to run
1364 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1368 run_test 55 "let MDS_CHECK_RESENT return the original return code instead of 0"
1370 #b3440 ASSERTION(rec->ur_fid2->id) failed
1372 ln -s foo $DIR/$tfile
1373 replay_barrier $SINGLEMDS
1374 #drop_reply "cat $DIR/$tfile"
1378 run_test 56 "don't replay a symlink open request (3440)"
1380 #recovery one mds-ost setattr from llog
1382 #define OBD_FAIL_MDS_OST_SETATTR 0x12c
1383 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012c"
1385 replay_barrier $SINGLEMDS
1388 $CHECKSTAT -t file $DIR/$tfile || return 1
1389 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1392 run_test 57 "test recovery from llog for setattr op"
1394 #recovery many mds-ost setattr from llog
1397 #define OBD_FAIL_MDS_OST_SETATTR 0x12c
1398 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012c"
1399 createmany -o $DIR/$tdir/$tfile-%d 2500
1400 replay_barrier $SINGLEMDS
1403 $CHECKSTAT -t file $DIR/$tdir/$tfile-* >/dev/null || return 1
1404 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1405 unlinkmany $DIR/$tdir/$tfile-%d 2500
1408 run_test 58a "test recovery from llog for setattr op (test llog_gen_rec)"
1414 large_xattr_enabled &&
1415 orig="$(generate_string $(max_xattr_size))" || orig="bar"
1417 mount_client $MOUNT2
1419 touch $DIR/$tdir/$tfile
1420 replay_barrier $SINGLEMDS
1421 setfattr -n trusted.foo -v $orig $DIR/$tdir/$tfile
1423 new=$(get_xattr_value trusted.foo $MOUNT2/$tdir/$tfile)
1424 [[ "$new" = "$orig" ]] || return 1
1425 rm -f $DIR/$tdir/$tfile
1427 zconf_umount `hostname` $MOUNT2
1429 run_test 58b "test replay of setxattr op"
1431 test_58c() { # bug 16570
1436 if large_xattr_enabled; then
1437 local xattr_size=$(max_xattr_size)
1438 orig="$(generate_string $((xattr_size / 2)))"
1439 orig1="$(generate_string $xattr_size)"
1445 mount_client $MOUNT2
1447 touch $DIR/$tdir/$tfile
1448 drop_request "setfattr -n trusted.foo -v $orig $DIR/$tdir/$tfile" ||
1450 new=$(get_xattr_value trusted.foo $MOUNT2/$tdir/$tfile)
1451 [[ "$new" = "$orig" ]] || return 2
1452 drop_reint_reply "setfattr -n trusted.foo1 -v $orig1 $DIR/$tdir/$tfile" ||
1454 new=$(get_xattr_value trusted.foo1 $MOUNT2/$tdir/$tfile)
1455 [[ "$new" = "$orig1" ]] || return 4
1456 rm -f $DIR/$tdir/$tfile
1458 zconf_umount $HOSTNAME $MOUNT2
1460 run_test 58c "resend/reconstruct setxattr op"
1462 # log_commit_thread vs filter_destroy race used to lead to import use after free
1465 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1468 createmany -o $DIR/$tdir/$tfile-%d 200
1470 unlinkmany $DIR/$tdir/$tfile-%d 200
1471 #define OBD_FAIL_PTLRPC_DELAY_RECOV 0x507
1472 do_facet ost1 "lctl set_param fail_loc=0x507"
1475 do_facet ost1 "lctl set_param fail_loc=0x0"
1479 run_test 59 "test log_commit_thread vs filter_destroy race"
1481 # race between add unlink llog vs cat log init in post_recovery (only for b1_6)
1482 # bug 12086: should no oops and No ctxt error for this test
1485 createmany -o $DIR/$tdir/$tfile-%d 200
1486 replay_barrier $SINGLEMDS
1487 unlinkmany $DIR/$tdir/$tfile-%d 0 100
1489 unlinkmany $DIR/$tdir/$tfile-%d 100 100
1490 local no_ctxt=`dmesg | grep "No ctxt"`
1491 [ -z "$no_ctxt" ] || error "ctxt is not initialized in recovery"
1493 run_test 60 "test llog post recovery init vs llog unlink"
1495 #test race llog recovery thread vs llog cleanup
1496 test_61a() { # was test_61
1497 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1500 createmany -o $DIR/$tdir/$tfile-%d 800
1502 # OBD_FAIL_OST_LLOG_RECOVERY_TIMEOUT 0x221
1503 unlinkmany $DIR/$tdir/$tfile-%d 800
1504 set_nodes_failloc "$(osts_nodes)" 0x80000221
1509 set_nodes_failloc "$(osts_nodes)" 0x0
1511 $CHECKSTAT -t file $DIR/$tdir/$tfile-* && return 1
1514 run_test 61a "test race llog recovery vs llog cleanup"
1516 #test race mds llog sync vs llog cleanup
1518 # OBD_FAIL_MDS_LLOG_SYNC_TIMEOUT 0x13a
1519 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013a"
1520 facet_failover $SINGLEMDS
1523 do_facet client dd if=/dev/zero of=$DIR/$tfile bs=4k count=1 || return 1
1525 run_test 61b "test race mds llog sync vs llog cleanup"
1527 #test race cancel cookie cb vs llog cleanup
1529 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1531 # OBD_FAIL_OST_CANCEL_COOKIE_TIMEOUT 0x222
1533 set_nodes_failloc "$(osts_nodes)" 0x80000222
1537 set_nodes_failloc "$(osts_nodes)" 0x0
1539 run_test 61c "test race mds llog sync vs llog cleanup"
1541 test_61d() { # bug 16002 # bug 17466 # bug 22137
1542 # OBD_FAIL_OBD_LLOG_SETUP 0x605
1544 do_facet mgs "lctl set_param fail_loc=0x80000605"
1545 start mgs $MGSDEV $MGS_MOUNT_OPTS && error "mgs start should have failed"
1546 do_facet mgs "lctl set_param fail_loc=0"
1547 start mgs $MGSDEV $MGS_MOUNT_OPTS || error "cannot restart mgs"
1549 run_test 61d "error in llog_setup should cleanup the llog context correctly"
1551 test_62() { # Bug 15756 - don't mis-drop resent replay
1553 replay_barrier $SINGLEMDS
1554 createmany -o $DIR/$tdir/$tfile- 25
1555 #define OBD_FAIL_TGT_REPLAY_DROP 0x707
1556 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000707"
1558 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1559 unlinkmany $DIR/$tdir/$tfile- 25 || return 2
1562 run_test 62 "don't mis-drop resent replay"
1564 #Adaptive Timeouts (bug 3055)
1572 echo "Cleaning up AT ..."
1573 if [ -n "$ATOLDBASE" ]; then
1574 local at_history=$($LCTL get_param -n at_history)
1575 do_facet $SINGLEMDS "lctl set_param at_history=$at_history" || true
1576 do_facet ost1 "lctl set_param at_history=$at_history" || true
1579 if [ $AT_MAX_SET -ne 0 ]; then
1580 for facet in mds client ost; do
1581 var=AT_MAX_SAVE_${facet}
1582 echo restore AT on $facet to saved value ${!var}
1583 at_max_set ${!var} $facet
1584 at_new=$(at_max_get $facet)
1585 echo Restored AT value on $facet $at_new
1586 [ $at_new -eq ${!var} ] || \
1587 error "$facet : AT value was not restored SAVED ${!var} NEW $at_new"
1594 local at_max_new=600
1596 # Save at_max original values
1598 if [ $AT_MAX_SET -eq 0 ]; then
1599 # Suppose that all osts have the same at_max
1600 for facet in mds client ost; do
1601 eval AT_MAX_SAVE_${facet}=$(at_max_get $facet)
1605 for facet in mds client ost; do
1606 at_max=$(at_max_get $facet)
1607 if [ $at_max -ne $at_max_new ]; then
1608 echo "AT value on $facet is $at_max, set it by force temporarily to $at_max_new"
1609 at_max_set $at_max_new $facet
1614 if [ -z "$ATOLDBASE" ]; then
1615 ATOLDBASE=$(do_facet $SINGLEMDS "lctl get_param -n at_history")
1616 # speed up the timebase so we can check decreasing AT
1617 do_facet $SINGLEMDS "lctl set_param at_history=8" || true
1618 do_facet ost1 "lctl set_param at_history=8" || true
1620 # sleep for a while to cool down, should be > 8s and also allow
1621 # at least one ping to be sent. simply use TIMEOUT to be safe.
1626 test_65a() #bug 3055
1628 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1630 at_start || return 0
1631 $LCTL dk > /dev/null
1633 $LCTL set_param debug="other"
1634 # Slow down a request to the current service time, this is critical
1635 # because previous tests may have caused this value to increase.
1636 REQ_DELAY=`lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts |
1637 awk '/portal 12/ {print $5}'`
1638 REQ_DELAY=$((${REQ_DELAY} + ${REQ_DELAY} / 4 + 5))
1640 do_facet $SINGLEMDS lctl set_param fail_val=$((${REQ_DELAY} * 1000))
1641 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1642 do_facet $SINGLEMDS $LCTL set_param fail_loc=0x8000050a
1643 createmany -o $DIR/$tfile 10 > /dev/null
1644 unlinkmany $DIR/$tfile 10 > /dev/null
1645 # check for log message
1646 $LCTL dk | grep "Early reply #" || error "No early reply"
1648 # client should show REQ_DELAY estimates
1649 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep portal
1651 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep portal
1653 run_test 65a "AT: verify early replies"
1655 test_65b() #bug 3055
1657 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1659 at_start || return 0
1662 $LCTL set_param debug="other trace"
1663 $LCTL dk > /dev/null
1664 # Slow down a request to the current service time, this is critical
1665 # because previous tests may have caused this value to increase.
1666 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tfile
1667 multiop $DIR/$tfile Ow1yc
1668 REQ_DELAY=`lctl get_param -n osc.${FSNAME}-OST0000-osc-*.timeouts |
1669 awk '/portal 6/ {print $5}'`
1670 REQ_DELAY=$((${REQ_DELAY} + ${REQ_DELAY} / 4 + 5))
1672 do_facet ost1 lctl set_param fail_val=${REQ_DELAY}
1673 #define OBD_FAIL_OST_BRW_PAUSE_PACK 0x224
1674 do_facet ost1 $LCTL set_param fail_loc=0x224
1677 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tfile
1678 # force some real bulk transfer
1679 multiop $DIR/$tfile oO_CREAT:O_RDWR:O_SYNC:w4096c
1681 do_facet ost1 $LCTL set_param fail_loc=0
1682 # check for log message
1683 $LCTL dk | grep "Early reply #" || error "No early reply"
1685 # client should show REQ_DELAY estimates
1686 lctl get_param -n osc.${FSNAME}-OST0000-osc-*.timeouts | grep portal
1688 run_test 65b "AT: verify early replies on packed reply / bulk"
1690 test_66a() #bug 3055
1692 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1694 at_start || return 0
1695 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1696 # adjust 5s at a time so no early reply is sent (within deadline)
1697 do_facet $SINGLEMDS "$LCTL set_param fail_val=5000"
1698 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1699 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0x8000050a"
1700 createmany -o $DIR/$tfile 20 > /dev/null
1701 unlinkmany $DIR/$tfile 20 > /dev/null
1702 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1703 do_facet $SINGLEMDS "$LCTL set_param fail_val=10000"
1704 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0x8000050a"
1705 createmany -o $DIR/$tfile 20 > /dev/null
1706 unlinkmany $DIR/$tfile 20 > /dev/null
1707 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1708 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0"
1710 createmany -o $DIR/$tfile 20 > /dev/null
1711 unlinkmany $DIR/$tfile 20 > /dev/null
1712 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1713 CUR=$(lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | awk '/portal 12/ {print $5}')
1714 WORST=$(lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | awk '/portal 12/ {print $7}')
1715 echo "Current MDT timeout $CUR, worst $WORST"
1716 [ $CUR -lt $WORST ] || error "Current $CUR should be less than worst $WORST"
1718 run_test 66a "AT: verify MDT service time adjusts with no early replies"
1720 test_66b() #bug 3055
1722 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1724 at_start || return 0
1725 ORIG=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1726 awk '/network/ {print $4}')
1727 $LCTL set_param fail_val=$(($ORIG + 5))
1728 #define OBD_FAIL_PTLRPC_PAUSE_REP 0x50c
1729 $LCTL set_param fail_loc=0x50c
1730 ls $DIR/$tfile > /dev/null 2>&1
1731 $LCTL set_param fail_loc=0
1732 CUR=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1733 awk '/network/ {print $4}')
1734 WORST=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1735 awk '/network/ {print $6}')
1736 echo "network timeout orig $ORIG, cur $CUR, worst $WORST"
1737 [ $WORST -gt $ORIG ] ||
1738 error "Worst $WORST should be worse than orig $ORIG"
1740 run_test 66b "AT: verify net latency adjusts"
1742 test_67a() #bug 3055
1744 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1746 at_start || return 0
1747 CONN1=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1748 # sleeping threads may drive values above this
1749 do_facet ost1 "$LCTL set_param fail_val=400"
1750 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1751 do_facet ost1 "$LCTL set_param fail_loc=0x50a"
1752 createmany -o $DIR/$tfile 20 > /dev/null
1753 unlinkmany $DIR/$tfile 20 > /dev/null
1754 do_facet ost1 "$LCTL set_param fail_loc=0"
1755 CONN2=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1756 ATTEMPTS=$(($CONN2 - $CONN1))
1757 echo "$ATTEMPTS osc reconnect attempts on gradual slow"
1758 [ $ATTEMPTS -gt 0 ] && error_ignore 13721 "AT should have prevented reconnect"
1761 run_test 67a "AT: verify slow request processing doesn't induce reconnects"
1763 test_67b() #bug 3055
1765 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1767 at_start || return 0
1768 CONN1=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1770 # exhaust precreations on ost1
1771 local OST=$(ostname_from_index 0)
1772 local mdtosc=$(get_mdtosc_proc_path mds $OST)
1773 local last_id=$(do_facet $SINGLEMDS lctl get_param -n \
1774 osc.$mdtosc.prealloc_last_id)
1775 local next_id=$(do_facet $SINGLEMDS lctl get_param -n \
1776 osc.$mdtosc.prealloc_next_id)
1778 mkdir -p $DIR/$tdir/${OST}
1779 $SETSTRIPE -i 0 -c 1 $DIR/$tdir/${OST} || error "$SETSTRIPE"
1780 echo "Creating to objid $last_id on ost $OST..."
1781 #define OBD_FAIL_OST_PAUSE_CREATE 0x223
1782 do_facet ost1 "$LCTL set_param fail_val=20000"
1783 do_facet ost1 "$LCTL set_param fail_loc=0x80000223"
1784 createmany -o $DIR/$tdir/${OST}/f $next_id $((last_id - next_id + 2))
1787 do_facet ost1 "lctl get_param -n ost.OSS.ost_create.timeouts"
1789 CONN2=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1790 ATTEMPTS=$(($CONN2 - $CONN1))
1791 echo "$ATTEMPTS osc reconnect attempts on instant slow"
1792 # do it again; should not timeout
1793 do_facet ost1 "$LCTL set_param fail_loc=0x80000223"
1794 cp /etc/profile $DIR/$tfile || error "cp failed"
1795 do_facet ost1 "$LCTL set_param fail_loc=0"
1797 do_facet ost1 "lctl get_param -n ost.OSS.ost_create.timeouts"
1798 CONN3=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1799 ATTEMPTS=$(($CONN3 - $CONN2))
1800 echo "$ATTEMPTS osc reconnect attempts on 2nd slow"
1801 [ $ATTEMPTS -gt 0 ] && error "AT should have prevented reconnect"
1804 run_test 67b "AT: verify instant slowdown doesn't induce reconnects"
1806 test_68 () #bug 13813
1808 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1810 at_start || return 0
1811 local ldlm_enqueue_min=$(find /sys -name ldlm_enqueue_min)
1812 [ -z "$ldlm_enqueue_min" ] && skip "missing /sys/.../ldlm_enqueue_min" && return 0
1813 local ldlm_enqueue_min_r=$(do_facet ost1 "find /sys -name ldlm_enqueue_min")
1814 [ -z "$ldlm_enqueue_min_r" ] && skip "missing /sys/.../ldlm_enqueue_min in the ost1" && return 0
1815 local ENQ_MIN=$(cat $ldlm_enqueue_min)
1816 local ENQ_MIN_R=$(do_facet ost1 "cat $ldlm_enqueue_min_r")
1817 echo $TIMEOUT >> $ldlm_enqueue_min
1818 do_facet ost1 "echo $TIMEOUT >> $ldlm_enqueue_min_r"
1822 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tdir
1823 #define OBD_FAIL_LDLM_PAUSE_CANCEL 0x312
1824 $LCTL set_param fail_val=$(($TIMEOUT - 1))
1825 $LCTL set_param fail_loc=0x80000312
1826 cp /etc/profile $DIR/$tdir/${tfile}_1 || error "1st cp failed $?"
1827 $LCTL set_param fail_val=$((TIMEOUT * 5 / 4))
1828 $LCTL set_param fail_loc=0x80000312
1829 cp /etc/profile $DIR/$tdir/${tfile}_2 || error "2nd cp failed $?"
1830 $LCTL set_param fail_loc=0
1832 echo $ENQ_MIN >> $ldlm_enqueue_min
1833 do_facet ost1 "echo $ENQ_MIN_R >> $ldlm_enqueue_min_r"
1837 run_test 68 "AT: verify slowing locks"
1840 # end of AT tests includes above lines
1843 # start multi-client tests
1845 [ -z "$CLIENTS" ] && \
1846 { skip "Need two or more clients." && return; }
1847 [ $CLIENTCOUNT -lt 2 ] && \
1848 { skip "Need two or more clients, have $CLIENTCOUNT" && return; }
1850 echo "mount clients $CLIENTS ..."
1851 zconf_mount_clients $CLIENTS $MOUNT
1853 local clients=${CLIENTS//,/ }
1854 echo "Write/read files on $DIR ; clients $CLIENTS ... "
1855 for CLIENT in $clients; do
1856 do_node $CLIENT dd bs=1M count=10 if=/dev/zero \
1857 of=$DIR/${tfile}_${CLIENT} 2>/dev/null || \
1858 error "dd failed on $CLIENT"
1861 local prev_client=$(echo $clients | sed 's/^.* \(.\+\)$/\1/')
1862 for C in ${CLIENTS//,/ }; do
1863 do_node $prev_client dd if=$DIR/${tfile}_${C} of=/dev/null 2>/dev/null || \
1864 error "dd if=$DIR/${tfile}_${C} failed on $prev_client"
1870 run_test 70a "check multi client t-f"
1872 check_for_process () {
1877 killall_process $clients "$prog" -0
1880 killall_process () {
1881 local clients=${1:-$(hostname)}
1886 do_nodes $clients "killall $signal $name"
1890 local clients=${CLIENTS:-$HOSTNAME}
1892 zconf_mount_clients $clients $MOUNT
1895 [ "$SLOW" = "no" ] && duration=120
1896 # set duration to 900 because it takes some time to boot node
1897 [ "$FAILURE_MODE" = HARD ] && duration=900
1900 local start_ts=$(date +%s)
1901 local cmd="rundbench 1 -t $duration"
1903 do_nodesv $clients "set -x; MISSING_DBENCH_OK=$MISSING_DBENCH_OK \
1904 PATH=\$PATH:$LUSTRE/utils:$LUSTRE/tests/:$DBENCH_LIB \
1905 DBENCH_LIB=$DBENCH_LIB TESTSUITE=$TESTSUITE TESTNAME=$TESTNAME \
1906 MOUNT=$MOUNT DIR=$DIR/$tdir/\\\$(hostname) LCTL=$LCTL $cmd" &
1909 #LU-1897 wait for all dbench copies to start
1910 while ! check_for_process $clients dbench; do
1911 elapsed=$(($(date +%s) - start_ts))
1912 if [ $elapsed -gt $duration]; then
1913 killall_process $clients dbench
1914 error "dbench failed to start on $clients!"
1919 log "Started rundbench load pid=$pid ..."
1921 elapsed=$(($(date +%s) - start_ts))
1922 local num_failovers=0
1923 while [ $elapsed -lt $duration ]; do
1924 if ! check_for_process $clients dbench; then
1925 error_noexit "dbench stopped on some of $clients!"
1926 killall_process $clients dbench
1930 replay_barrier $SINGLEMDS
1931 sleep 1 # give clients a time to do operations
1932 # Increment the number of failovers
1933 num_failovers=$((num_failovers+1))
1934 log "$TESTNAME fail $SINGLEMDS $num_failovers times"
1936 elapsed=$(($(date +%s) - start_ts))
1939 wait $pid || error "rundbench load on $clients failed!"
1941 run_test 70b "mds recovery; $CLIENTCOUNT clients"
1942 # end multi-client tests
1945 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1949 replay_barrier $SINGLEMDS
1950 #define OBD_FAIL_LDLM_ENQUEUE_NET 0x302
1951 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000302"
1954 wait $pid || return 1
1955 [ -e $DIR/$tfile ] && return 2
1958 run_test 73a "open(O_CREAT), unlink, replay, reconnect before open replay , close"
1961 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1965 replay_barrier $SINGLEMDS
1966 #define OBD_FAIL_LDLM_REPLY 0x30c
1967 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000030c"
1970 wait $pid || return 1
1971 [ -e $DIR/$tfile ] && return 2
1974 run_test 73b "open(O_CREAT), unlink, replay, reconnect at open_replay reply, close"
1977 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1981 replay_barrier $SINGLEMDS
1982 #define OBD_FAIL_TGT_LAST_REPLAY 0x710
1983 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000710"
1986 wait $pid || return 1
1987 [ -e $DIR/$tfile ] && return 2
1990 run_test 73c "open(O_CREAT), unlink, replay, reconnect at last_replay, close"
1994 local clients=${CLIENTS:-$HOSTNAME}
1996 zconf_umount_clients $clients $MOUNT
1998 facet_failover $SINGLEMDS
1999 zconf_mount_clients $clients $MOUNT
2001 touch $DIR/$tfile || return 1
2002 rm $DIR/$tfile || return 2
2003 clients_up || error "client evicted: $?"
2006 run_test 74 "Ensure applications don't fail waiting for OST recovery"
2008 remote_dir_check_80() {
2010 local diridx=$($GETSTRIPE -M $remote_dir)
2011 [ $diridx -eq $MDTIDX ] || error "$diridx != $MDTIDX"
2013 createmany -o $remote_dir/f-%d 20 || error "creation failed"
2014 local fileidx=$($GETSTRIPE -M $remote_dir/f-1)
2015 [ $fileidx -eq $MDTIDX ] || error "$fileidx != $MDTIDX"
2021 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2022 ([ $FAILURE_MODE == "HARD" ] &&
2023 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2024 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2028 local remote_dir=$DIR/$tdir/remote_dir
2031 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2032 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1701
2033 $LFS mkdir -i $MDTIDX $remote_dir &
2036 fail mds$((MDTIDX + 1))
2038 wait $CLIENT_PID || error "remote creation failed"
2040 remote_dir_check_80 || error "remote dir check failed"
2041 rm -rf $DIR/$tdir || error "rmdir failed"
2045 run_test 80a "DNE: create remote dir, drop update rep from MDT1, fail MDT1"
2048 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2049 ([ $FAILURE_MODE == "HARD" ] &&
2050 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2051 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2055 local remote_dir=$DIR/$tdir/remote_dir
2058 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2059 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1701
2060 $LFS mkdir -i $MDTIDX $remote_dir &
2065 wait $CLIENT_PID || error "remote creation failed"
2067 remote_dir_check_80 || error "remote dir check failed"
2068 rm -rf $DIR/$tdir || error "rmdir failed"
2072 run_test 80b "DNE: create remote dir, drop update rep from MDT1, fail MDT0"
2075 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2076 ([ $FAILURE_MODE == "HARD" ] &&
2077 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2078 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2082 local remote_dir=$DIR/$tdir/remote_dir
2085 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2086 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1701
2087 $LFS mkdir -i $MDTIDX $remote_dir &
2091 fail mds$((MDTIDX + 1))
2093 wait $CLIENT_PID || error "remote creation failed"
2095 remote_dir_check_80 || error "remote dir check failed"
2096 rm -rf $DIR/$tdir || error "rmdir failed"
2100 run_test 80c "DNE: create remote dir, drop update rep from MDT1, fail MDT[0,1]"
2103 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2105 local remote_dir=$DIR/$tdir/remote_dir
2108 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2109 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1701
2110 $LFS mkdir -i $MDTIDX $remote_dir &
2113 fail mds${MDTIDX},mds$((MDTIDX + 1))
2115 wait $CLIENT_PID || error "remote creation failed"
2117 remote_dir_check_80 || error "remote dir check failed"
2118 rm -rf $DIR/$tdir || error "rmdir failed"
2122 run_test 80d "DNE: create remote dir, drop update rep from MDT1, fail 2 MDTs"
2125 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2126 ([ $FAILURE_MODE == "HARD" ] &&
2127 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2128 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2132 local remote_dir=$DIR/$tdir/remote_dir
2135 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2136 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2137 $LFS mkdir -i $MDTIDX $remote_dir &
2142 wait $CLIENT_PID || error "remote creation failed"
2144 remote_dir_check_80 || error "remote dir check failed"
2145 rm -rf $DIR/$tdir || error "rmdir failed"
2149 run_test 80e "DNE: create remote dir, drop MDT0 rep, fail MDT0"
2152 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2153 ([ $FAILURE_MODE == "HARD" ] &&
2154 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2155 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2158 local remote_dir=$DIR/$tdir/remote_dir
2161 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2162 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2163 $LFS mkdir -i $MDTIDX $remote_dir &
2166 fail mds$((MDTIDX + 1))
2168 wait $CLIENT_PID || error "remote creation failed"
2170 remote_dir_check_80 || error "remote dir check failed"
2171 rm -rf $DIR/$tdir || error "rmdir failed"
2175 run_test 80f "DNE: create remote dir, drop MDT0 rep, fail MDT1"
2178 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2179 ([ $FAILURE_MODE == "HARD" ] &&
2180 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2181 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2185 local remote_dir=$DIR/$tdir/remote_dir
2188 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2189 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2190 $LFS mkdir -i $MDTIDX $remote_dir &
2194 fail mds$((MDTIDX + 1))
2196 wait $CLIENT_PID || error "remote creation failed"
2198 remote_dir_check_80 || error "remote dir check failed"
2199 rm -rf $DIR/$tdir || error "rmdir failed"
2203 run_test 80g "DNE: create remote dir, drop MDT0 rep, fail MDT0, then MDT1"
2206 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2208 local remote_dir=$DIR/$tdir/remote_dir
2211 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2212 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2213 $LFS mkdir -i $MDTIDX $remote_dir &
2216 fail mds${MDTIDX},mds$((MDTIDX + 1))
2218 wait $CLIENT_PID || return 1
2220 remote_dir_check_80 || error "remote dir check failed"
2221 rm -rf $DIR/$tdir || error "rmdir failed"
2225 run_test 80h "DNE: create remote dir, drop MDT0 rep, fail 2 MDTs"
2228 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2229 ([ $FAILURE_MODE == "HARD" ] &&
2230 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2231 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2235 local remote_dir=$DIR/$tdir/remote_dir
2238 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2241 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2242 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2246 fail mds$((MDTIDX + 1))
2248 wait $CLIENT_PID || error "rm remote dir failed"
2250 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2252 rm -rf $DIR/$tdir || error "rmdir failed"
2256 run_test 81a "DNE: unlink remote dir, drop MDT0 update rep, fail MDT1"
2259 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2260 ([ $FAILURE_MODE == "HARD" ] &&
2261 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2262 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2265 local remote_dir=$DIR/$tdir/remote_dir
2268 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2270 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2271 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2277 wait $CLIENT_PID || error "rm remote dir failed"
2279 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2281 rm -rf $DIR/$tdir || error "rmdir failed"
2285 run_test 81b "DNE: unlink remote dir, drop MDT0 update reply, fail MDT0"
2288 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2289 ([ $FAILURE_MODE == "HARD" ] &&
2290 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2291 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2295 local remote_dir=$DIR/$tdir/remote_dir
2298 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2300 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2301 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2306 fail mds$((MDTIDX + 1))
2308 wait $CLIENT_PID || error "rm remote dir failed"
2310 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2312 rm -rf $DIR/$tdir || error "rmdir failed"
2316 run_test 81c "DNE: unlink remote dir, drop MDT0 update reply, fail MDT0,MDT1"
2319 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2321 local remote_dir=$DIR/$tdir/remote_dir
2324 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2326 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2327 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2331 fail mds${MDTIDX},mds$((MDTIDX + 1))
2333 wait $CLIENT_PID || error "rm remote dir failed"
2335 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2337 rm -rf $DIR/$tdir || error "rmdir failed"
2341 run_test 81d "DNE: unlink remote dir, drop MDT0 update reply, fail 2 MDTs"
2344 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2345 ([ $FAILURE_MODE == "HARD" ] &&
2346 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2347 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2351 local remote_dir=$DIR/$tdir/remote_dir
2354 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2356 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2357 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2360 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0
2364 wait $CLIENT_PID || error "rm remote dir failed"
2366 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2368 rm -rf $DIR/$tdir || error "rmdir failed"
2372 run_test 81e "DNE: unlink remote dir, drop MDT1 req reply, fail MDT0"
2375 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2376 ([ $FAILURE_MODE == "HARD" ] &&
2377 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2378 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2382 local remote_dir=$DIR/$tdir/remote_dir
2385 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2387 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2388 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2392 fail mds$((MDTIDX + 1))
2394 wait $CLIENT_PID || error "rm remote dir failed"
2396 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2398 rm -rf $DIR/$tdir || error "rmdir failed"
2402 run_test 81f "DNE: unlink remote dir, drop MDT1 req reply, fail MDT1"
2405 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2406 ([ $FAILURE_MODE == "HARD" ] &&
2407 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2408 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2412 local remote_dir=$DIR/$tdir/remote_dir
2415 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2417 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2418 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2423 fail mds$((MDTIDX + 1))
2425 wait $CLIENT_PID || error "rm remote dir failed"
2427 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2429 rm -rf $DIR/$tdir || error "rmdir failed"
2433 run_test 81g "DNE: unlink remote dir, drop req reply, fail M0, then M1"
2436 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2438 local remote_dir=$DIR/$tdir/remote_dir
2441 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2443 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2444 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2448 fail mds${MDTIDX},mds$((MDTIDX + 1))
2450 wait $CLIENT_PID || error "rm remote dir failed"
2452 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2454 rm -rf $DIR/$tdir || error "rmdir failed"
2458 run_test 81h "DNE: unlink remote dir, drop request reply, fail 2 MDTs"
2462 createmany -o $DIR/$tdir/$tfile- 10 || return 1
2463 #define OBD_FAIL_MDS_FAIL_LOV_LOG_ADD 0x140
2464 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000140"
2465 unlinkmany $DIR/$tdir/$tfile- 10 || return 2
2467 run_test 83a "fail log_add during unlink recovery"
2471 createmany -o $DIR/$tdir/$tfile- 10 || return 1
2472 replay_barrier $SINGLEMDS
2473 unlinkmany $DIR/$tdir/$tfile- 10 || return 2
2474 #define OBD_FAIL_MDS_FAIL_LOV_LOG_ADD 0x140
2475 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000140"
2478 run_test 83b "fail log_add during unlink recovery"
2481 #define OBD_FAIL_MDS_OPEN_WAIT_CREATE 0x144
2482 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000144"
2483 createmany -o $DIR/$tfile- 1 &
2487 client_up || client_up || true # reconnect
2489 run_test 84a "stale open during export disconnect"
2491 test_85a() { #bug 16774
2492 lctl set_param -n ldlm.cancel_unused_locks_before_replay "1"
2494 for i in `seq 100`; do
2495 echo "tag-$i" > $DIR/$tfile-$i
2496 grep -q "tag-$i" $DIR/$tfile-$i || error "f2-$i"
2499 lov_id=`lctl dl | grep "clilov"`
2500 addr=`echo $lov_id | awk '{print $4}' | awk -F '-' '{print $3}'`
2501 count=`lctl get_param -n ldlm.namespaces.*MDT0000*$addr.lock_unused_count`
2502 echo "before recovery: unused locks count = $count"
2506 count2=`lctl get_param -n ldlm.namespaces.*MDT0000*$addr.lock_unused_count`
2507 echo "after recovery: unused locks count = $count2"
2509 if [ $count2 -ge $count ]; then
2510 error "unused locks are not canceled"
2513 run_test 85a "check the cancellation of unused locks during recovery(IBITS)"
2515 test_85b() { #bug 16774
2516 lctl set_param -n ldlm.cancel_unused_locks_before_replay "1"
2518 do_facet mgs $LCTL pool_new $FSNAME.$TESTNAME || return 1
2519 do_facet mgs $LCTL pool_add $FSNAME.$TESTNAME $FSNAME-OST0000 || return 2
2521 $SETSTRIPE -c 1 -p $FSNAME.$TESTNAME $DIR
2523 for i in `seq 100`; do
2524 dd if=/dev/urandom of=$DIR/$tfile-$i bs=4096 count=32 >/dev/null 2>&1
2527 cancel_lru_locks osc
2529 for i in `seq 100`; do
2530 dd if=$DIR/$tfile-$i of=/dev/null bs=4096 count=32 >/dev/null 2>&1
2533 lov_id=`lctl dl | grep "clilov"`
2534 addr=`echo $lov_id | awk '{print $4}' | awk -F '-' '{print $3}'`
2535 count=`lctl get_param -n ldlm.namespaces.*OST0000*$addr.lock_unused_count`
2536 echo "before recovery: unused locks count = $count"
2537 [ $count != 0 ] || return 3
2541 count2=`lctl get_param -n ldlm.namespaces.*OST0000*$addr.lock_unused_count`
2542 echo "after recovery: unused locks count = $count2"
2544 do_facet mgs $LCTL pool_remove $FSNAME.$TESTNAME $FSNAME-OST0000 || return 4
2545 do_facet mgs $LCTL pool_destroy $FSNAME.$TESTNAME || return 5
2547 if [ $count2 -ge $count ]; then
2548 error "unused locks are not canceled"
2551 run_test 85b "check the cancellation of unused locks during recovery(EXTENT)"
2554 local clients=${CLIENTS:-$HOSTNAME}
2556 zconf_umount_clients $clients $MOUNT
2557 do_facet $SINGLEMDS lctl set_param mdt.${FSNAME}-MDT*.exports.clear=0
2558 remount_facet $SINGLEMDS
2559 zconf_mount_clients $clients $MOUNT
2561 run_test 86 "umount server after clear nid_stats should not hit LBUG"
2564 do_facet ost1 "lctl set_param -n obdfilter.${ost1_svc}.sync_journal 0"
2567 $SETSTRIPE -i 0 -c 1 $DIR/$tfile
2568 dd if=/dev/urandom of=$DIR/$tfile bs=1024k count=8 || error "Cannot write"
2569 cksum=`md5sum $DIR/$tfile | awk '{print $1}'`
2570 cancel_lru_locks osc
2572 dd if=$DIR/$tfile of=/dev/null bs=1024k count=8 || error "Cannot read"
2573 cksum2=`md5sum $DIR/$tfile | awk '{print $1}'`
2574 if [ $cksum != $cksum2 ] ; then
2575 error "New checksum $cksum2 does not match original $cksum"
2578 run_test 87 "write replay"
2581 do_facet ost1 "lctl set_param -n obdfilter.${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 sleep 1 # Give it a chance to flush dirty data
2587 echo TESTTEST | dd of=$DIR/$tfile bs=1 count=8 seek=64
2588 cksum=`md5sum $DIR/$tfile | awk '{print $1}'`
2589 cancel_lru_locks osc
2591 dd if=$DIR/$tfile of=/dev/null bs=1024k count=8 || error "Cannot read"
2592 cksum2=`md5sum $DIR/$tfile | awk '{print $1}'`
2593 if [ $cksum != $cksum2 ] ; then
2594 error "New checksum $cksum2 does not match original $cksum"
2597 run_test 87b "write replay with changed data (checksum resend)"
2599 test_88() { #bug 17485
2603 $SETSTRIPE -i 0 -c 1 $DIR/$tdir || error "$SETSTRIPE"
2606 replay_barrier $SINGLEMDS
2608 # exhaust precreations on ost1
2609 local OST=$(ostname_from_index 0)
2610 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $OST)
2611 local last_id=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_last_id)
2612 local next_id=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_next_id)
2613 echo "before test: last_id = $last_id, next_id = $next_id"
2615 echo "Creating to objid $last_id on ost $OST..."
2616 createmany -o $DIR/$tdir/f-%d $next_id $((last_id - next_id + 2))
2618 #create some files to use some uncommitted objids
2619 last_id=$(($last_id + 1))
2620 createmany -o $DIR/$tdir/f-%d $last_id 8
2622 last_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_last_id)
2623 next_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_next_id)
2624 echo "before recovery: last_id = $last_id2, next_id = $next_id2"
2626 # if test uses shutdown_facet && reboot_facet instead of facet_failover ()
2627 # it has to take care about the affected facets, bug20407
2628 local affected_mds1=$(affected_facets mds1)
2629 local affected_ost1=$(affected_facets ost1)
2631 shutdown_facet $SINGLEMDS
2634 reboot_facet $SINGLEMDS
2635 change_active $affected_mds1
2636 wait_for_facet $affected_mds1
2637 mount_facets $affected_mds1 || error "Restart of mds failed"
2640 change_active $affected_ost1
2641 wait_for_facet $affected_ost1
2642 mount_facets $affected_ost1 || error "Restart of ost1 failed"
2646 last_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_last_id)
2647 next_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_next_id)
2648 echo "after recovery: last_id = $last_id2, next_id = $next_id2"
2650 # create new files, which should use new objids, and ensure the orphan
2651 # cleanup phase for ost1 is completed at the same time
2652 for i in `seq 8`; do
2653 file_id=$(($last_id + 10 + $i))
2654 dd if=/dev/urandom of=$DIR/$tdir/f-$file_id bs=4096 count=128
2657 # if the objids were not recreated, then "ls" will failed for -ENOENT
2658 ls -l $DIR/$tdir/* || error "can't get the status of precreated files"
2661 # write into previously created files
2662 for i in `seq 8`; do
2663 file_id=$(($last_id + $i))
2664 dd if=/dev/urandom of=$DIR/$tdir/f-$file_id bs=4096 count=128
2665 cp -f $DIR/$tdir/f-$file_id $TMP/$tdir/
2668 # compare the content
2669 for i in `seq 8`; do
2670 file_id=$(($last_id + $i))
2671 cmp $TMP/$tdir/f-$file_id $DIR/$tdir/f-$file_id || error "the content" \
2672 "of file is modified!"
2677 run_test 88 "MDS should not assign same objid to different files "
2680 cancel_lru_locks osc
2682 rm -f $DIR/$tdir/$tfile
2684 wait_delete_completed
2685 BLOCKS1=$(df -P $MOUNT | tail -n 1 | awk '{ print $3 }')
2686 $SETSTRIPE -i 0 -c 1 $DIR/$tdir/$tfile
2687 dd if=/dev/zero bs=1M count=10 of=$DIR/$tdir/$tfile
2690 facet_failover $SINGLEMDS
2691 rm $DIR/$tdir/$tfile
2694 zconf_mount $(hostname) $MOUNT
2695 client_up || return 1
2697 wait_delete_completed
2698 BLOCKS2=$(df -P $MOUNT | tail -n 1 | awk '{ print $3 }')
2699 [ $((BLOCKS2 - BLOCKS1)) -le 4 ] || \
2700 error $((BLOCKS2 - BLOCKS1)) blocks leaked
2703 run_test 89 "no disk space leak on late ost connection"
2709 change_active $facet
2710 wait_for_facet $facet
2711 mount_facet $facet || error "Restart of $facet failed"
2715 test_90() { # bug 19494
2716 local dir=$DIR/$tdir
2717 local ostfail=$(get_random_entry $(get_facets OST))
2719 if [[ $FAILURE_MODE = HARD ]]; then
2720 local affected=$(affected_facets $ostfail);
2721 if [[ "$affected" != $ostfail ]]; then
2722 skip not functional with FAILURE_MODE=$FAILURE_MODE, affected: $affected
2729 echo "Create the files"
2731 # file "f${index}" striped over 1 OST
2732 # file "all" striped over all OSTs
2734 $SETSTRIPE -c $OSTCOUNT $dir/all ||
2735 error "setstripe failed to create $dir/all"
2737 for (( i=0; i<$OSTCOUNT; i++ )); do
2739 $SETSTRIPE -i $i -c 1 $f || error "$SETSTRIPE failed to create $f"
2741 # confirm setstripe actually created the stripe on the requested OST
2742 local uuid=$(ostuuid_from_index $i)
2743 for file in f$i all; do
2744 if [[ $dir/$file != $($LFS find --obd $uuid --name $file $dir) ]]; then
2745 $GETSTRIPE $dir/file
2746 error wrong stripe: $file, uuid: $uuid
2751 # Before failing an OST, get its obd name and index
2752 local varsvc=${ostfail}_svc
2753 local obd=$(do_facet $ostfail lctl get_param -n obdfilter.${!varsvc}.uuid)
2754 local index=$(($(facet_number $ostfail) - 1))
2756 echo "Fail $ostfail $obd, display the list of affected files"
2757 shutdown_facet $ostfail || return 2
2759 trap "cleanup_90 $ostfail" EXIT INT
2760 echo "General Query: lfs find $dir"
2761 local list=$($LFS find $dir)
2763 for (( i=0; i<$OSTCOUNT; i++ )); do
2764 list_member "$list" $dir/f$i || error_noexit "lfs find $dir: no file f$i"
2766 list_member "$list" $dir/all || error_noexit "lfs find $dir: no file all"
2768 # focus on the missing OST,
2769 # we expect to see only two files affected: "f$(index)" and "all"
2771 echo "Querying files on shutdown $ostfail: lfs find --obd $obd"
2772 list=$($LFS find --obd $obd $dir)
2774 for file in all f$index; do
2775 list_member "$list" $dir/$file ||
2776 error_noexit "lfs find does not report the affected $obd for $file"
2779 [[ $(echo $list | wc -w) -eq 2 ]] ||
2780 error_noexit "lfs find reports the wrong list of affected files ${#list[@]}"
2782 echo "Check getstripe: $GETSTRIPE -r --obd $obd"
2783 list=$($GETSTRIPE -r --obd $obd $dir)
2785 for file in all f$index; do
2786 echo "$list" | grep $dir/$file ||
2787 error_noexit "lfs getsripe does not report the affected $obd for $file"
2792 run_test 90 "lfs find identifies the missing striped file segments"
2795 check_and_cleanup_lustre