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 fail $SINGLEMDS # start orphan recovery
444 wait_recovery_complete $SINGLEMDS || error "MDS recovery not done"
445 wait_delete_completed_mds $wait_timeout || return 3
446 AFTERUSED=$(df -P $DIR | tail -1 | awk '{ print $3 }')
447 log "before $BEFOREUSED, after $AFTERUSED"
448 (( $AFTERUSED > $BEFOREUSED + $(fs_log_size) )) &&
449 error "after $AFTERUSED > before $BEFOREUSED"
452 run_test 20b "write, unlink, eviction, replay, (test mds_cleanup_orphans)"
454 test_20c() { # bug 10480
455 multiop_bg_pause $DIR/$tfile Ow_c || return 1
461 client_up || client_up || true # reconnect
464 wait $pid || return 1
465 [ -s $DIR/$tfile ] || error "File was truncated"
469 run_test 20c "check that client eviction does not affect file content"
472 replay_barrier $SINGLEMDS
473 multiop_bg_pause $DIR/$tfile O_tSc || return 5
476 touch $DIR/g11 || return 1
480 wait $pid || return 2
481 [ -e $DIR/$tfile ] && return 3
482 touch $DIR/h11 || return 4
485 run_test 21 "|X| open(O_CREAT), unlink touch new, replay, close (test mds_cleanup_orphans)"
488 multiop_bg_pause $DIR/$tfile O_tSc || return 3
491 replay_barrier $SINGLEMDS
496 wait $pid || return 1
497 [ -e $DIR/$tfile ] && return 2
500 run_test 22 "open(O_CREAT), |X| unlink, replay, close (test mds_cleanup_orphans)"
503 multiop_bg_pause $DIR/$tfile O_tSc || return 5
506 replay_barrier $SINGLEMDS
508 touch $DIR/g11 || return 1
512 wait $pid || return 2
513 [ -e $DIR/$tfile ] && return 3
514 touch $DIR/h11 || return 4
517 run_test 23 "open(O_CREAT), |X| unlink touch new, replay, close (test mds_cleanup_orphans)"
520 multiop_bg_pause $DIR/$tfile O_tSc || return 3
523 replay_barrier $SINGLEMDS
527 wait $pid || return 1
528 [ -e $DIR/$tfile ] && return 2
531 run_test 24 "open(O_CREAT), replay, unlink, close (test mds_cleanup_orphans)"
534 multiop_bg_pause $DIR/$tfile O_tSc || return 3
538 replay_barrier $SINGLEMDS
541 wait $pid || return 1
542 [ -e $DIR/$tfile ] && return 2
545 run_test 25 "open(O_CREAT), unlink, replay, close (test mds_cleanup_orphans)"
548 replay_barrier $SINGLEMDS
549 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
551 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
556 wait $pid2 || return 1
560 wait $pid1 || return 2
561 [ -e $DIR/$tfile-1 ] && return 3
562 [ -e $DIR/$tfile-2 ] && return 4
565 run_test 26 "|X| open(O_CREAT), unlink two, close one, replay, close one (test mds_cleanup_orphans)"
568 replay_barrier $SINGLEMDS
569 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
571 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
578 wait $pid1 || return 1
580 wait $pid2 || return 2
581 [ -e $DIR/$tfile-1 ] && return 3
582 [ -e $DIR/$tfile-2 ] && return 4
585 run_test 27 "|X| open(O_CREAT), unlink two, replay, close two (test mds_cleanup_orphans)"
588 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
590 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
592 replay_barrier $SINGLEMDS
596 wait $pid2 || return 1
600 wait $pid1 || return 2
601 [ -e $DIR/$tfile-1 ] && return 3
602 [ -e $DIR/$tfile-2 ] && return 4
605 run_test 28 "open(O_CREAT), |X| unlink two, close one, replay, close one (test mds_cleanup_orphans)"
608 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
610 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
612 replay_barrier $SINGLEMDS
618 wait $pid1 || return 1
620 wait $pid2 || return 2
621 [ -e $DIR/$tfile-1 ] && return 3
622 [ -e $DIR/$tfile-2 ] && return 4
625 run_test 29 "open(O_CREAT), |X| unlink two, replay, close two (test mds_cleanup_orphans)"
628 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
630 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
635 replay_barrier $SINGLEMDS
638 wait $pid1 || return 1
640 wait $pid2 || return 2
641 [ -e $DIR/$tfile-1 ] && return 3
642 [ -e $DIR/$tfile-2 ] && return 4
645 run_test 30 "open(O_CREAT) two, unlink two, replay, close two (test mds_cleanup_orphans)"
648 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
650 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
654 replay_barrier $SINGLEMDS
658 wait $pid1 || return 1
660 wait $pid2 || return 2
661 [ -e $DIR/$tfile-1 ] && return 3
662 [ -e $DIR/$tfile-2 ] && return 4
665 run_test 31 "open(O_CREAT) two, unlink one, |X| unlink one, close two (test mds_cleanup_orphans)"
667 # tests for bug 2104; completion without crashing is success. The close is
668 # stale, but we always return 0 for close, so the app never sees it.
670 multiop_bg_pause $DIR/$tfile O_c || return 2
672 multiop_bg_pause $DIR/$tfile O_c || return 3
675 client_up || client_up || return 1
678 wait $pid1 || return 4
679 wait $pid2 || return 5
682 run_test 32 "close() notices client eviction; close() after client eviction"
685 createmany -o $DIR/$tfile-%d 10
686 replay_barrier_nosync $SINGLEMDS
687 fail_abort $SINGLEMDS
688 # recreate shouldn't fail
689 createmany -o $DIR/$tfile--%d 10 || return 1
693 run_test 33a "fid seq shouldn't be reused after abort recovery"
696 #define OBD_FAIL_SEQ_ALLOC 0x1311
697 do_facet $SINGLEMDS "lctl set_param fail_loc=0x1311"
699 createmany -o $DIR/$tfile-%d 10
700 replay_barrier_nosync $SINGLEMDS
701 fail_abort $SINGLEMDS
702 # recreate shouldn't fail
703 createmany -o $DIR/$tfile--%d 10 || return 1
707 run_test 33b "test fid seq allocation"
710 multiop_bg_pause $DIR/$tfile O_c || return 2
714 replay_barrier $SINGLEMDS
715 fail_abort $SINGLEMDS
717 wait $pid || return 3
718 [ -e $DIR/$tfile ] && return 1
722 run_test 34 "abort recovery before client does replay (test mds_cleanup_orphans)"
724 # bug 2278 - generate one orphan on OST, then destroy it during recovery from llog
728 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
729 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000119"
734 # give a chance to remove from MDS
735 fail_abort $SINGLEMDS
736 $CHECKSTAT -t file $DIR/$tfile && return 1 || true
738 run_test 35 "test recovery from llog for unlink op"
740 # b=2432 resent cancel after replay uses wrong cookie,
741 # so don't resend cancels
743 replay_barrier $SINGLEMDS
745 checkstat $DIR/$tfile
746 facet_failover $SINGLEMDS
748 if dmesg | grep "unknown lock cookie"; then
749 echo "cancel after replay failed"
753 run_test 36 "don't resend cancel"
756 # directory orphans can't be unlinked from PENDING directory
758 rmdir $DIR/$tfile 2>/dev/null
759 multiop_bg_pause $DIR/$tfile dD_c || return 2
763 replay_barrier $SINGLEMDS
764 # clear the dmesg buffer so we only see errors from this recovery
765 do_facet $SINGLEMDS dmesg -c >/dev/null
766 fail_abort $SINGLEMDS
768 do_facet $SINGLEMDS dmesg | grep "error .* unlinking .* from PENDING" &&
770 wait $pid || return 3
774 start_full_debug_logging
775 run_test 37 "abort recovery before client does replay (test mds_cleanup_orphans for directories)"
776 stop_full_debug_logging
779 createmany -o $DIR/$tfile-%d 800
780 unlinkmany $DIR/$tfile-%d 0 400
781 replay_barrier $SINGLEMDS
783 unlinkmany $DIR/$tfile-%d 400 400
785 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
787 run_test 38 "test recovery from unlink llog (test llog_gen_rec) "
789 test_39() { # bug 4176
790 createmany -o $DIR/$tfile-%d 800
791 replay_barrier $SINGLEMDS
792 unlinkmany $DIR/$tfile-%d 0 400
794 unlinkmany $DIR/$tfile-%d 400 400
796 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
798 run_test 39 "test recovery from unlink llog (test llog_gen_rec) "
801 lctl get_param -n osc.*.stats | awk -vwrites=0 '/ost_write/ { writes += $2 } END { print writes; }'
806 # always need connection to MDS to verify layout during IO. LU-2628.
807 lctl get_param mdc.*.connect_flags | grep -q layout_lock &&
808 skip "layout_lock needs MDS connection for IO" && return 0
810 $LCTL mark multiop $MOUNT/$tfile OS_c
811 multiop $MOUNT/$tfile OS_c &
813 writeme -s $MOUNT/${tfile}-2 &
816 facet_failover $SINGLEMDS
817 #define OBD_FAIL_MDS_CONNECT_NET 0x117
818 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000117"
820 stat1=`count_ost_writes`
822 stat2=`count_ost_writes`
823 echo "$stat1, $stat2"
824 if [ $stat1 -lt $stat2 ]; then
825 echo "writes continuing during recovery"
828 echo "writes not continuing during recovery, bug 2477"
831 echo "waiting for writeme $WRITE_PID"
835 echo "waiting for multiop $PID"
836 wait $PID || return 2
837 do_facet client munlink $MOUNT/$tfile || return 3
838 do_facet client munlink $MOUNT/${tfile}-2 || return 3
841 run_test 40 "cause recovery in ptlrpc, ensure IO continues"
845 # make sure that a read to one osc doesn't try to double-unlock its page just
846 # because another osc is invalid. trigger_group_io used to mistakenly return
847 # an error if any oscs were invalid even after having successfully put rpcs
848 # on valid oscs. This was fatal if the caller was ll_readpage who unlocked
849 # the page, guarnateeing that the unlock from the RPC completion would
850 # assert on trying to unlock the unlocked page.
852 [ $OSTCOUNT -lt 2 ] &&
853 skip_env "skipping test 41: we don't have a second OST to test with" &&
856 local f=$MOUNT/$tfile
857 # make sure the start of the file is ost1
858 $SETSTRIPE -S $((128 * 1024)) -i 0 $f
859 do_facet client dd if=/dev/zero of=$f bs=4k count=1 || return 3
861 # fail ost2 and read from ost1
862 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $ost2_svc)
863 local osc2dev=$(do_facet $SINGLEMDS "lctl get_param -n devices" | \
864 grep $mdtosc | awk '{print $1}')
865 [ -z "$osc2dev" ] && echo "OST: $ost2_svc" && lctl get_param -n devices &&
867 do_facet $SINGLEMDS $LCTL --device $osc2dev deactivate || return 1
868 do_facet client dd if=$f of=/dev/null bs=4k count=1 || return 3
869 do_facet $SINGLEMDS $LCTL --device $osc2dev activate || return 2
872 run_test 41 "read from a valid osc while other oscs are invalid"
874 # test MDS recovery after ost failure
876 blocks=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
877 createmany -o $DIR/$tfile-%d 800
879 unlinkmany $DIR/$tfile-%d 0 400
881 lctl set_param debug=-1
884 # osc is evicted, fs is smaller (but only with failout OSTs (bug 7287)
885 #blocks_after=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
886 #[ $blocks_after -lt $blocks ] || return 1
887 echo wait for MDS to timeout and recover
888 sleep $((TIMEOUT * 2))
890 unlinkmany $DIR/$tfile-%d 400 400
891 $CHECKSTAT -t file $DIR/$tfile-* && return 2 || true
893 run_test 42 "recovery after ost failure"
895 # timeout in MDS/OST recovery RPC will LBUG MDS
896 test_43() { # bug 2530
897 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
899 replay_barrier $SINGLEMDS
901 # OBD_FAIL_OST_CREATE_NET 0x204
902 do_facet ost1 "lctl set_param fail_loc=0x80000204"
905 do_facet ost1 "lctl set_param fail_loc=0"
909 run_test 43 "mds osc import failure during recovery; don't LBUG"
911 test_44a() { # was test_44
914 local mdcdev=$($LCTL dl |
915 awk "/${FSNAME}-MDT0000-mdc-/ {if (\$2 == \"UP\") {print \$1}}")
916 [ "$mdcdev" ] || return 2
917 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
918 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
920 # adaptive timeouts slow this way down
921 if at_is_enabled; then
922 at_max_saved=$(at_max_get mds)
926 for i in `seq 1 10`; do
927 echo "$i of 10 ($(date +%s))"
928 do_facet $SINGLEMDS \
929 "lctl get_param -n md[ts].*.mdt.timeouts | grep service"
930 #define OBD_FAIL_TGT_CONN_RACE 0x701
931 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000701"
932 # lctl below may fail, it is valid case
933 $LCTL --device $mdcdev recover
936 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
937 [ $at_max_saved -ne 0 ] && at_max_set $at_max_saved mds
940 run_test 44a "race in target handle connect"
943 local mdcdev=$($LCTL dl |
944 awk "/${FSNAME}-MDT0000-mdc-/ {if (\$2 == \"UP\") {print \$1}}")
945 [ "$mdcdev" ] || return 2
946 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
947 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
949 for i in `seq 1 10`; do
950 echo "$i of 10 ($(date +%s))"
951 do_facet $SINGLEMDS \
952 "lctl get_param -n md[ts].*.mdt.timeouts | grep service"
953 #define OBD_FAIL_TGT_DELAY_RECONNECT 0x704
954 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000704"
955 # lctl below may fail, it is valid case
956 $LCTL --device $mdcdev recover
959 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
962 run_test 44b "race in target handle connect"
965 replay_barrier $SINGLEMDS
966 createmany -m $DIR/$tfile-%d 100 || error "failed to create directories"
967 #define OBD_FAIL_TGT_RCVG_FLAG 0x712
968 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000712"
969 fail_abort $SINGLEMDS
970 unlinkmany $DIR/$tfile-%d 100 && error "unliked after fail abort"
972 unlinkmany $DIR/$tfile-%d 100 && error "unliked after fail"
975 run_test 44c "race in target handle connect"
977 # Handle failed close
979 local mdcdev=$($LCTL get_param -n devices |
980 awk "/ ${FSNAME}-MDT0000-mdc-/ {print \$1}")
981 [ "$mdcdev" ] || return 2
982 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
983 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
985 $LCTL --device $mdcdev recover || return 6
987 multiop_bg_pause $DIR/$tfile O_c || return 1
990 # This will cause the CLOSE to fail before even
991 # allocating a reply buffer
992 $LCTL --device $mdcdev deactivate || return 4
996 wait $pid || return 1
998 $LCTL --device $mdcdev activate || return 5
1001 $CHECKSTAT -t file $DIR/$tfile || return 2
1004 run_test 45 "Handle failed close"
1008 drop_reply "touch $DIR/$tfile"
1010 # ironically, the previous test, 45, will cause a real forced close,
1011 # so just look for one for this test
1012 dmesg | grep -i "force closing client file handle for $tfile" && return 1
1015 run_test 46 "Don't leak file handle after open resend (3325)"
1017 test_47() { # bug 2824
1018 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1020 # create some files to make sure precreate has been done on all
1021 # OSTs. (just in case this test is run independently)
1022 createmany -o $DIR/$tfile 20 || return 1
1024 # OBD_FAIL_OST_CREATE_NET 0x204
1026 do_facet ost1 "lctl set_param fail_loc=0x80000204"
1027 client_up || return 2
1029 # let the MDS discover the OST failure, attempt to recover, fail
1030 # and recover again.
1031 sleep $((3 * TIMEOUT))
1033 # Without 2824, this createmany would hang
1034 createmany -o $DIR/$tfile 20 || return 3
1035 unlinkmany $DIR/$tfile 20 || return 4
1037 do_facet ost1 "lctl set_param fail_loc=0"
1040 run_test 47 "MDS->OSC failure during precreate cleanup (2824)"
1043 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1044 [ "$OSTCOUNT" -lt "2" ] && skip_env "$OSTCOUNT < 2 OSTs -- skipping" && return
1046 replay_barrier $SINGLEMDS
1047 createmany -o $DIR/$tfile 20 || return 1
1048 # OBD_FAIL_OST_EROFS 0x216
1049 facet_failover $SINGLEMDS
1050 do_facet ost1 "lctl set_param fail_loc=0x80000216"
1051 client_up || return 2
1053 createmany -o $DIR/$tfile 20 20 || return 2
1054 unlinkmany $DIR/$tfile 40 || return 3
1057 run_test 48 "MDS->OSC failure during precreate cleanup (2824)"
1060 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $ost1_svc)
1061 local oscdev=$(do_facet $SINGLEMDS "lctl get_param -n devices" | \
1062 grep $mdtosc | awk '{print $1}')
1063 [ "$oscdev" ] || return 1
1064 do_facet $SINGLEMDS $LCTL --device $oscdev recover || return 2
1065 do_facet $SINGLEMDS $LCTL --device $oscdev recover || return 3
1066 # give the mds_lov_sync threads a chance to run
1069 run_test 50 "Double OSC recovery, don't LASSERT (3812)"
1071 # b3764 timed out lock replay
1074 cancel_lru_locks mdc
1076 multiop $DIR/$tfile s || return 1
1077 replay_barrier $SINGLEMDS
1078 #define OBD_FAIL_LDLM_REPLY 0x30c
1079 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000030c"
1080 fail $SINGLEMDS || return 2
1081 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1083 $CHECKSTAT -t file $DIR/$tfile-* && return 3 || true
1085 run_test 52 "time out lock replay (3764)"
1087 # bug 3462 - simultaneous MDC requests
1089 cancel_lru_locks mdc # cleanup locks from former test cases
1090 mkdir -p $DIR/${tdir}-1
1091 mkdir -p $DIR/${tdir}-2
1092 multiop $DIR/${tdir}-1/f O_c &
1094 # give multiop a change to open
1097 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1098 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1099 kill -USR1 $close_pid
1100 cancel_lru_locks mdc # force the close
1101 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1103 mcreate $DIR/${tdir}-2/f || return 1
1105 # close should still be here
1106 [ -d /proc/$close_pid ] || return 2
1108 replay_barrier_nodf $SINGLEMDS
1110 wait $close_pid || return 3
1112 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1113 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1114 rm -rf $DIR/${tdir}-*
1116 run_test 53a "|X| close request while two MDC requests in flight"
1119 cancel_lru_locks mdc # cleanup locks from former test cases
1120 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1122 mkdir -p $DIR/${tdir}-1
1123 mkdir -p $DIR/${tdir}-2
1124 multiop_bg_pause $DIR/${tdir}-1/f O_c || return 6
1127 #define OBD_FAIL_MDS_REINT_NET 0x107
1128 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1129 mcreate $DIR/${tdir}-2/f &
1133 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1134 kill -USR1 $close_pid
1135 cancel_lru_locks mdc # force the close
1136 wait $close_pid || return 1
1137 # open should still be here
1138 [ -d /proc/$open_pid ] || return 2
1140 replay_barrier_nodf $SINGLEMDS
1142 wait $open_pid || return 3
1144 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1145 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1146 rm -rf $DIR/${tdir}-*
1148 run_test 53b "|X| open request while two MDC requests in flight"
1151 cancel_lru_locks mdc # cleanup locks from former test cases
1152 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1154 mkdir -p $DIR/${tdir}-1
1155 mkdir -p $DIR/${tdir}-2
1156 multiop $DIR/${tdir}-1/f O_c &
1159 #define OBD_FAIL_MDS_REINT_NET 0x107
1160 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1161 mcreate $DIR/${tdir}-2/f &
1165 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1166 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1167 kill -USR1 $close_pid
1168 cancel_lru_locks mdc # force the close
1170 #bz20647: make sure all pids are exists before failover
1171 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1172 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1173 replay_barrier_nodf $SINGLEMDS
1174 fail_nodf $SINGLEMDS
1175 wait $open_pid || return 1
1177 # close should be gone
1178 [ -d /proc/$close_pid ] && return 2
1179 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1181 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1182 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1183 rm -rf $DIR/${tdir}-*
1185 run_test 53c "|X| open request and close request while two MDC requests in flight"
1188 cancel_lru_locks mdc # cleanup locks from former test cases
1189 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1191 mkdir -p $DIR/${tdir}-1
1192 mkdir -p $DIR/${tdir}-2
1193 multiop $DIR/${tdir}-1/f O_c &
1195 # give multiop a chance to open
1198 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1199 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1200 kill -USR1 $close_pid
1201 cancel_lru_locks mdc # force the close
1202 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1203 mcreate $DIR/${tdir}-2/f || return 1
1205 # close should still be here
1206 [ -d /proc/$close_pid ] || return 2
1208 wait $close_pid || return 3
1210 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1211 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1212 rm -rf $DIR/${tdir}-*
1214 run_test 53d "|X| close reply while two MDC requests in flight"
1217 cancel_lru_locks mdc # cleanup locks from former test cases
1218 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1220 mkdir -p $DIR/${tdir}-1
1221 mkdir -p $DIR/${tdir}-2
1222 multiop $DIR/${tdir}-1/f O_c &
1225 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1226 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1227 mcreate $DIR/${tdir}-2/f &
1231 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1232 kill -USR1 $close_pid
1233 cancel_lru_locks mdc # force the close
1234 wait $close_pid || return 1
1235 # open should still be here
1236 [ -d /proc/$open_pid ] || return 2
1238 replay_barrier_nodf $SINGLEMDS
1240 wait $open_pid || return 3
1242 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1243 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1244 rm -rf $DIR/${tdir}-*
1246 run_test 53e "|X| open reply while two MDC requests in flight"
1249 cancel_lru_locks mdc # cleanup locks from former test cases
1250 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1252 mkdir -p $DIR/${tdir}-1
1253 mkdir -p $DIR/${tdir}-2
1254 multiop $DIR/${tdir}-1/f O_c &
1257 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1258 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1259 mcreate $DIR/${tdir}-2/f &
1263 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1264 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1265 kill -USR1 $close_pid
1266 cancel_lru_locks mdc # force the close
1268 #bz20647: make sure all pids are exists before failover
1269 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1270 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1271 replay_barrier_nodf $SINGLEMDS
1272 fail_nodf $SINGLEMDS
1273 wait $open_pid || return 1
1275 # close should be gone
1276 [ -d /proc/$close_pid ] && return 2
1277 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1279 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1280 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1281 rm -rf $DIR/${tdir}-*
1283 run_test 53f "|X| open reply and close reply while two MDC requests in flight"
1286 cancel_lru_locks mdc # cleanup locks from former test cases
1287 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1289 mkdir -p $DIR/${tdir}-1
1290 mkdir -p $DIR/${tdir}-2
1291 multiop $DIR/${tdir}-1/f O_c &
1294 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1295 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1296 mcreate $DIR/${tdir}-2/f &
1300 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1301 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1302 kill -USR1 $close_pid
1303 cancel_lru_locks mdc # force the close
1304 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1306 #bz20647: make sure all pids are exists before failover
1307 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1308 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1309 replay_barrier_nodf $SINGLEMDS
1310 fail_nodf $SINGLEMDS
1311 wait $open_pid || return 1
1313 # close should be gone
1314 [ -d /proc/$close_pid ] && return 2
1316 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1317 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1318 rm -rf $DIR/${tdir}-*
1320 run_test 53g "|X| drop open reply and close request while close and open are both in flight"
1323 cancel_lru_locks mdc # cleanup locks from former test cases
1324 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1326 mkdir -p $DIR/${tdir}-1
1327 mkdir -p $DIR/${tdir}-2
1328 multiop $DIR/${tdir}-1/f O_c &
1331 #define OBD_FAIL_MDS_REINT_NET 0x107
1332 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1333 mcreate $DIR/${tdir}-2/f &
1337 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1338 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1339 kill -USR1 $close_pid
1340 cancel_lru_locks mdc # force the close
1343 #bz20647: make sure all pids are exists before failover
1344 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1345 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1346 replay_barrier_nodf $SINGLEMDS
1347 fail_nodf $SINGLEMDS
1348 wait $open_pid || return 1
1350 # close should be gone
1351 [ -d /proc/$close_pid ] && return 2
1352 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1354 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1355 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1356 rm -rf $DIR/${tdir}-*
1358 run_test 53h "|X| open request and close reply while two MDC requests in flight"
1360 #b_cray 54 "|X| open request and close reply while two MDC requests in flight"
1362 #b3761 ASSERTION(hash != 0) failed
1364 # OBD_FAIL_MDS_OPEN_CREATE | OBD_FAIL_ONCE
1365 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012b"
1367 # give touch a chance to run
1369 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1373 run_test 55 "let MDS_CHECK_RESENT return the original return code instead of 0"
1375 #b3440 ASSERTION(rec->ur_fid2->id) failed
1377 ln -s foo $DIR/$tfile
1378 replay_barrier $SINGLEMDS
1379 #drop_reply "cat $DIR/$tfile"
1383 run_test 56 "don't replay a symlink open request (3440)"
1385 #recovery one mds-ost setattr from llog
1387 #define OBD_FAIL_MDS_OST_SETATTR 0x12c
1388 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012c"
1390 replay_barrier $SINGLEMDS
1393 $CHECKSTAT -t file $DIR/$tfile || return 1
1394 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1397 run_test 57 "test recovery from llog for setattr op"
1400 zconf_umount `hostname` $MOUNT2
1404 #recovery many mds-ost setattr from llog
1407 #define OBD_FAIL_MDS_OST_SETATTR 0x12c
1408 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012c"
1409 createmany -o $DIR/$tdir/$tfile-%d 2500
1410 replay_barrier $SINGLEMDS
1413 $CHECKSTAT -t file $DIR/$tdir/$tfile-* >/dev/null || return 1
1414 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1415 unlinkmany $DIR/$tdir/$tfile-%d 2500
1418 run_test 58a "test recovery from llog for setattr op (test llog_gen_rec)"
1424 trap cleanup_58 EXIT
1426 large_xattr_enabled &&
1427 orig="$(generate_string $(max_xattr_size))" || orig="bar"
1429 mount_client $MOUNT2
1431 touch $DIR/$tdir/$tfile
1432 replay_barrier $SINGLEMDS
1433 setfattr -n trusted.foo -v $orig $DIR/$tdir/$tfile
1435 new=$(get_xattr_value trusted.foo $MOUNT2/$tdir/$tfile)
1436 [[ "$new" = "$orig" ]] || return 1
1437 rm -f $DIR/$tdir/$tfile
1441 run_test 58b "test replay of setxattr op"
1443 test_58c() { # bug 16570
1448 trap cleanup_58 EXIT
1450 if large_xattr_enabled; then
1451 local xattr_size=$(max_xattr_size)
1452 orig="$(generate_string $((xattr_size / 2)))"
1453 orig1="$(generate_string $xattr_size)"
1459 mount_client $MOUNT2
1461 touch $DIR/$tdir/$tfile
1462 drop_request "setfattr -n trusted.foo -v $orig $DIR/$tdir/$tfile" ||
1464 new=$(get_xattr_value trusted.foo $MOUNT2/$tdir/$tfile)
1465 [[ "$new" = "$orig" ]] || return 2
1466 drop_reint_reply "setfattr -n trusted.foo1 -v $orig1 $DIR/$tdir/$tfile" ||
1468 new=$(get_xattr_value trusted.foo1 $MOUNT2/$tdir/$tfile)
1469 [[ "$new" = "$orig1" ]] || return 4
1470 rm -f $DIR/$tdir/$tfile
1474 run_test 58c "resend/reconstruct setxattr op"
1476 # log_commit_thread vs filter_destroy race used to lead to import use after free
1479 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1482 createmany -o $DIR/$tdir/$tfile-%d 200
1484 unlinkmany $DIR/$tdir/$tfile-%d 200
1485 #define OBD_FAIL_PTLRPC_DELAY_RECOV 0x507
1486 do_facet ost1 "lctl set_param fail_loc=0x507"
1489 do_facet ost1 "lctl set_param fail_loc=0x0"
1493 run_test 59 "test log_commit_thread vs filter_destroy race"
1495 # race between add unlink llog vs cat log init in post_recovery (only for b1_6)
1496 # bug 12086: should no oops and No ctxt error for this test
1499 createmany -o $DIR/$tdir/$tfile-%d 200
1500 replay_barrier $SINGLEMDS
1501 unlinkmany $DIR/$tdir/$tfile-%d 0 100
1503 unlinkmany $DIR/$tdir/$tfile-%d 100 100
1504 local no_ctxt=`dmesg | grep "No ctxt"`
1505 [ -z "$no_ctxt" ] || error "ctxt is not initialized in recovery"
1507 run_test 60 "test llog post recovery init vs llog unlink"
1509 #test race llog recovery thread vs llog cleanup
1510 test_61a() { # was test_61
1511 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1514 createmany -o $DIR/$tdir/$tfile-%d 800
1516 # OBD_FAIL_OST_LLOG_RECOVERY_TIMEOUT 0x221
1517 unlinkmany $DIR/$tdir/$tfile-%d 800
1518 set_nodes_failloc "$(osts_nodes)" 0x80000221
1523 set_nodes_failloc "$(osts_nodes)" 0x0
1525 $CHECKSTAT -t file $DIR/$tdir/$tfile-* && return 1
1528 run_test 61a "test race llog recovery vs llog cleanup"
1530 #test race mds llog sync vs llog cleanup
1532 # OBD_FAIL_MDS_LLOG_SYNC_TIMEOUT 0x13a
1533 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013a"
1534 facet_failover $SINGLEMDS
1537 do_facet client dd if=/dev/zero of=$DIR/$tfile bs=4k count=1 || return 1
1539 run_test 61b "test race mds llog sync vs llog cleanup"
1541 #test race cancel cookie cb vs llog cleanup
1543 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1545 # OBD_FAIL_OST_CANCEL_COOKIE_TIMEOUT 0x222
1547 set_nodes_failloc "$(osts_nodes)" 0x80000222
1551 set_nodes_failloc "$(osts_nodes)" 0x0
1553 run_test 61c "test race mds llog sync vs llog cleanup"
1555 test_61d() { # bug 16002 # bug 17466 # bug 22137
1556 # OBD_FAIL_OBD_LLOG_SETUP 0x605
1558 do_facet mgs "lctl set_param fail_loc=0x80000605"
1559 start mgs $(mgsdevname) $MGS_MOUNT_OPTS &&
1560 error "mgs start should have failed"
1561 do_facet mgs "lctl set_param fail_loc=0"
1562 start mgs $(mgsdevname) $MGS_MOUNT_OPTS || error "cannot restart mgs"
1564 run_test 61d "error in llog_setup should cleanup the llog context correctly"
1566 test_62() { # Bug 15756 - don't mis-drop resent replay
1568 replay_barrier $SINGLEMDS
1569 createmany -o $DIR/$tdir/$tfile- 25
1570 #define OBD_FAIL_TGT_REPLAY_DROP 0x707
1571 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000707"
1573 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1574 unlinkmany $DIR/$tdir/$tfile- 25 || return 2
1577 run_test 62 "don't mis-drop resent replay"
1579 #Adaptive Timeouts (bug 3055)
1587 echo "Cleaning up AT ..."
1588 if [ -n "$ATOLDBASE" ]; then
1589 local at_history=$($LCTL get_param -n at_history)
1590 do_facet $SINGLEMDS "lctl set_param at_history=$at_history" || true
1591 do_facet ost1 "lctl set_param at_history=$at_history" || true
1594 if [ $AT_MAX_SET -ne 0 ]; then
1595 for facet in mds client ost; do
1596 var=AT_MAX_SAVE_${facet}
1597 echo restore AT on $facet to saved value ${!var}
1598 at_max_set ${!var} $facet
1599 at_new=$(at_max_get $facet)
1600 echo Restored AT value on $facet $at_new
1601 [ $at_new -eq ${!var} ] || \
1602 error "$facet : AT value was not restored SAVED ${!var} NEW $at_new"
1609 local at_max_new=600
1611 # Save at_max original values
1613 if [ $AT_MAX_SET -eq 0 ]; then
1614 # Suppose that all osts have the same at_max
1615 for facet in mds client ost; do
1616 eval AT_MAX_SAVE_${facet}=$(at_max_get $facet)
1620 for facet in mds client ost; do
1621 at_max=$(at_max_get $facet)
1622 if [ $at_max -ne $at_max_new ]; then
1623 echo "AT value on $facet is $at_max, set it by force temporarily to $at_max_new"
1624 at_max_set $at_max_new $facet
1629 if [ -z "$ATOLDBASE" ]; then
1630 ATOLDBASE=$(do_facet $SINGLEMDS "lctl get_param -n at_history")
1631 # speed up the timebase so we can check decreasing AT
1632 do_facet $SINGLEMDS "lctl set_param at_history=8" || true
1633 do_facet ost1 "lctl set_param at_history=8" || true
1635 # sleep for a while to cool down, should be > 8s and also allow
1636 # at least one ping to be sent. simply use TIMEOUT to be safe.
1641 test_65a() #bug 3055
1643 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1645 at_start || return 0
1646 $LCTL dk > /dev/null
1648 $LCTL set_param debug="other"
1649 # Slow down a request to the current service time, this is critical
1650 # because previous tests may have caused this value to increase.
1651 REQ_DELAY=`lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts |
1652 awk '/portal 12/ {print $5}'`
1653 REQ_DELAY=$((${REQ_DELAY} + ${REQ_DELAY} / 4 + 5))
1655 do_facet $SINGLEMDS lctl set_param fail_val=$((${REQ_DELAY} * 1000))
1656 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1657 do_facet $SINGLEMDS $LCTL set_param fail_loc=0x8000050a
1658 createmany -o $DIR/$tfile 10 > /dev/null
1659 unlinkmany $DIR/$tfile 10 > /dev/null
1660 # check for log message
1661 $LCTL dk | grep "Early reply #" || error "No early reply"
1663 # client should show REQ_DELAY estimates
1664 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep portal
1666 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep portal
1668 run_test 65a "AT: verify early replies"
1670 test_65b() #bug 3055
1672 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1674 at_start || return 0
1677 $LCTL set_param debug="other trace"
1678 $LCTL dk > /dev/null
1679 # Slow down a request to the current service time, this is critical
1680 # because previous tests may have caused this value to increase.
1681 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tfile
1682 multiop $DIR/$tfile Ow1yc
1683 REQ_DELAY=`lctl get_param -n osc.${FSNAME}-OST0000-osc-*.timeouts |
1684 awk '/portal 6/ {print $5}'`
1685 REQ_DELAY=$((${REQ_DELAY} + ${REQ_DELAY} / 4 + 5))
1687 do_facet ost1 lctl set_param fail_val=${REQ_DELAY}
1688 #define OBD_FAIL_OST_BRW_PAUSE_PACK 0x224
1689 do_facet ost1 $LCTL set_param fail_loc=0x224
1692 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tfile
1693 # force some real bulk transfer
1694 multiop $DIR/$tfile oO_CREAT:O_RDWR:O_SYNC:w4096c
1696 do_facet ost1 $LCTL set_param fail_loc=0
1697 # check for log message
1698 $LCTL dk | grep "Early reply #" || error "No early reply"
1700 # client should show REQ_DELAY estimates
1701 lctl get_param -n osc.${FSNAME}-OST0000-osc-*.timeouts | grep portal
1703 run_test 65b "AT: verify early replies on packed reply / bulk"
1705 test_66a() #bug 3055
1707 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1709 at_start || return 0
1710 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1711 # adjust 5s at a time so no early reply is sent (within deadline)
1712 do_facet $SINGLEMDS "$LCTL set_param fail_val=5000"
1713 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1714 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0x8000050a"
1715 createmany -o $DIR/$tfile 20 > /dev/null
1716 unlinkmany $DIR/$tfile 20 > /dev/null
1717 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1718 do_facet $SINGLEMDS "$LCTL set_param fail_val=10000"
1719 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0x8000050a"
1720 createmany -o $DIR/$tfile 20 > /dev/null
1721 unlinkmany $DIR/$tfile 20 > /dev/null
1722 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1723 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0"
1725 createmany -o $DIR/$tfile 20 > /dev/null
1726 unlinkmany $DIR/$tfile 20 > /dev/null
1727 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1728 CUR=$(lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | awk '/portal 12/ {print $5}')
1729 WORST=$(lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | awk '/portal 12/ {print $7}')
1730 echo "Current MDT timeout $CUR, worst $WORST"
1731 [ $CUR -lt $WORST ] || error "Current $CUR should be less than worst $WORST"
1733 run_test 66a "AT: verify MDT service time adjusts with no early replies"
1735 test_66b() #bug 3055
1737 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1739 at_start || return 0
1740 ORIG=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1741 awk '/network/ {print $4}')
1742 $LCTL set_param fail_val=$(($ORIG + 5))
1743 #define OBD_FAIL_PTLRPC_PAUSE_REP 0x50c
1744 $LCTL set_param fail_loc=0x50c
1745 ls $DIR/$tfile > /dev/null 2>&1
1746 $LCTL set_param fail_loc=0
1747 CUR=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1748 awk '/network/ {print $4}')
1749 WORST=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1750 awk '/network/ {print $6}')
1751 echo "network timeout orig $ORIG, cur $CUR, worst $WORST"
1752 [ $WORST -gt $ORIG ] ||
1753 error "Worst $WORST should be worse than orig $ORIG"
1755 run_test 66b "AT: verify net latency adjusts"
1757 test_67a() #bug 3055
1759 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1761 at_start || return 0
1762 CONN1=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1763 # sleeping threads may drive values above this
1764 do_facet ost1 "$LCTL set_param fail_val=400"
1765 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1766 do_facet ost1 "$LCTL set_param fail_loc=0x50a"
1767 createmany -o $DIR/$tfile 20 > /dev/null
1768 unlinkmany $DIR/$tfile 20 > /dev/null
1769 do_facet ost1 "$LCTL set_param fail_loc=0"
1770 CONN2=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1771 ATTEMPTS=$(($CONN2 - $CONN1))
1772 echo "$ATTEMPTS osc reconnect attempts on gradual slow"
1773 [ $ATTEMPTS -gt 0 ] &&
1774 error_ignore bz13721 "AT should have prevented reconnect"
1777 run_test 67a "AT: verify slow request processing doesn't induce reconnects"
1779 test_67b() #bug 3055
1781 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1783 at_start || return 0
1784 CONN1=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1786 # exhaust precreations on ost1
1787 local OST=$(ostname_from_index 0)
1788 local mdtosc=$(get_mdtosc_proc_path mds $OST)
1789 local last_id=$(do_facet $SINGLEMDS lctl get_param -n \
1790 osc.$mdtosc.prealloc_last_id)
1791 local next_id=$(do_facet $SINGLEMDS lctl get_param -n \
1792 osc.$mdtosc.prealloc_next_id)
1794 mkdir -p $DIR/$tdir/${OST}
1795 $SETSTRIPE -i 0 -c 1 $DIR/$tdir/${OST} || error "$SETSTRIPE"
1796 echo "Creating to objid $last_id on ost $OST..."
1797 #define OBD_FAIL_OST_PAUSE_CREATE 0x223
1798 do_facet ost1 "$LCTL set_param fail_val=20000"
1799 do_facet ost1 "$LCTL set_param fail_loc=0x80000223"
1800 createmany -o $DIR/$tdir/${OST}/f $next_id $((last_id - next_id + 2))
1803 do_facet ost1 "lctl get_param -n ost.OSS.ost_create.timeouts"
1805 CONN2=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1806 ATTEMPTS=$(($CONN2 - $CONN1))
1807 echo "$ATTEMPTS osc reconnect attempts on instant slow"
1808 # do it again; should not timeout
1809 do_facet ost1 "$LCTL set_param fail_loc=0x80000223"
1810 cp /etc/profile $DIR/$tfile || error "cp failed"
1811 do_facet ost1 "$LCTL set_param fail_loc=0"
1813 do_facet ost1 "lctl get_param -n ost.OSS.ost_create.timeouts"
1814 CONN3=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1815 ATTEMPTS=$(($CONN3 - $CONN2))
1816 echo "$ATTEMPTS osc reconnect attempts on 2nd slow"
1817 [ $ATTEMPTS -gt 0 ] && error "AT should have prevented reconnect"
1820 run_test 67b "AT: verify instant slowdown doesn't induce reconnects"
1822 test_68 () #bug 13813
1824 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1826 at_start || return 0
1827 local ldlm_enqueue_min=$(find /sys -name ldlm_enqueue_min)
1828 [ -z "$ldlm_enqueue_min" ] && skip "missing /sys/.../ldlm_enqueue_min" && return 0
1829 local ldlm_enqueue_min_r=$(do_facet ost1 "find /sys -name ldlm_enqueue_min")
1830 [ -z "$ldlm_enqueue_min_r" ] && skip "missing /sys/.../ldlm_enqueue_min in the ost1" && return 0
1831 local ENQ_MIN=$(cat $ldlm_enqueue_min)
1832 local ENQ_MIN_R=$(do_facet ost1 "cat $ldlm_enqueue_min_r")
1833 echo $TIMEOUT >> $ldlm_enqueue_min
1834 do_facet ost1 "echo $TIMEOUT >> $ldlm_enqueue_min_r"
1838 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tdir
1839 #define OBD_FAIL_LDLM_PAUSE_CANCEL 0x312
1840 $LCTL set_param fail_val=$(($TIMEOUT - 1))
1841 $LCTL set_param fail_loc=0x80000312
1842 cp /etc/profile $DIR/$tdir/${tfile}_1 || error "1st cp failed $?"
1843 $LCTL set_param fail_val=$((TIMEOUT * 5 / 4))
1844 $LCTL set_param fail_loc=0x80000312
1845 cp /etc/profile $DIR/$tdir/${tfile}_2 || error "2nd cp failed $?"
1846 $LCTL set_param fail_loc=0
1848 echo $ENQ_MIN >> $ldlm_enqueue_min
1849 do_facet ost1 "echo $ENQ_MIN_R >> $ldlm_enqueue_min_r"
1853 run_test 68 "AT: verify slowing locks"
1856 # end of AT tests includes above lines
1859 # start multi-client tests
1861 [ -z "$CLIENTS" ] && \
1862 { skip "Need two or more clients." && return; }
1863 [ $CLIENTCOUNT -lt 2 ] && \
1864 { skip "Need two or more clients, have $CLIENTCOUNT" && return; }
1866 echo "mount clients $CLIENTS ..."
1867 zconf_mount_clients $CLIENTS $MOUNT
1869 local clients=${CLIENTS//,/ }
1870 echo "Write/read files on $DIR ; clients $CLIENTS ... "
1871 for CLIENT in $clients; do
1872 do_node $CLIENT dd bs=1M count=10 if=/dev/zero \
1873 of=$DIR/${tfile}_${CLIENT} 2>/dev/null || \
1874 error "dd failed on $CLIENT"
1877 local prev_client=$(echo $clients | sed 's/^.* \(.\+\)$/\1/')
1878 for C in ${CLIENTS//,/ }; do
1879 do_node $prev_client dd if=$DIR/${tfile}_${C} of=/dev/null 2>/dev/null || \
1880 error "dd if=$DIR/${tfile}_${C} failed on $prev_client"
1886 run_test 70a "check multi client t-f"
1888 check_for_process () {
1893 killall_process $clients "$prog" -0
1896 killall_process () {
1897 local clients=${1:-$(hostname)}
1902 do_nodes $clients "killall $signal $name"
1906 local clients=${CLIENTS:-$HOSTNAME}
1908 zconf_mount_clients $clients $MOUNT
1911 [ "$SLOW" = "no" ] && duration=120
1912 # set duration to 900 because it takes some time to boot node
1913 [ "$FAILURE_MODE" = HARD ] && duration=900
1916 local start_ts=$(date +%s)
1917 local cmd="rundbench 1 -t $duration"
1919 do_nodesv $clients "set -x; MISSING_DBENCH_OK=$MISSING_DBENCH_OK \
1920 PATH=\$PATH:$LUSTRE/utils:$LUSTRE/tests/:$DBENCH_LIB \
1921 DBENCH_LIB=$DBENCH_LIB TESTSUITE=$TESTSUITE TESTNAME=$TESTNAME \
1922 MOUNT=$MOUNT DIR=$DIR/$tdir/\\\$(hostname) LCTL=$LCTL $cmd" &
1925 #LU-1897 wait for all dbench copies to start
1926 while ! check_for_process $clients dbench; do
1927 elapsed=$(($(date +%s) - start_ts))
1928 if [ $elapsed -gt $duration ]; then
1929 killall_process $clients dbench
1930 error "dbench failed to start on $clients!"
1935 log "Started rundbench load pid=$pid ..."
1937 elapsed=$(($(date +%s) - start_ts))
1938 local num_failovers=0
1939 while [ $elapsed -lt $duration ]; do
1940 if ! check_for_process $clients dbench; then
1941 error_noexit "dbench stopped on some of $clients!"
1942 killall_process $clients dbench
1946 replay_barrier $SINGLEMDS
1947 sleep 1 # give clients a time to do operations
1948 # Increment the number of failovers
1949 num_failovers=$((num_failovers+1))
1950 log "$TESTNAME fail $SINGLEMDS $num_failovers times"
1952 elapsed=$(($(date +%s) - start_ts))
1955 wait $pid || error "rundbench load on $clients failed!"
1957 run_test 70b "mds recovery; $CLIENTCOUNT clients"
1958 # end multi-client tests
1961 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1965 replay_barrier $SINGLEMDS
1966 #define OBD_FAIL_LDLM_ENQUEUE_NET 0x302
1967 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000302"
1970 wait $pid || return 1
1971 [ -e $DIR/$tfile ] && return 2
1974 run_test 73a "open(O_CREAT), unlink, replay, reconnect before open replay , close"
1977 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1981 replay_barrier $SINGLEMDS
1982 #define OBD_FAIL_LDLM_REPLY 0x30c
1983 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000030c"
1986 wait $pid || return 1
1987 [ -e $DIR/$tfile ] && return 2
1990 run_test 73b "open(O_CREAT), unlink, replay, reconnect at open_replay reply, close"
1993 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1997 replay_barrier $SINGLEMDS
1998 #define OBD_FAIL_TGT_LAST_REPLAY 0x710
1999 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000710"
2002 wait $pid || return 1
2003 [ -e $DIR/$tfile ] && return 2
2006 run_test 73c "open(O_CREAT), unlink, replay, reconnect at last_replay, close"
2010 local clients=${CLIENTS:-$HOSTNAME}
2012 zconf_umount_clients $clients $MOUNT
2014 facet_failover $SINGLEMDS
2015 zconf_mount_clients $clients $MOUNT
2017 touch $DIR/$tfile || return 1
2018 rm $DIR/$tfile || return 2
2019 clients_up || error "client evicted: $?"
2022 run_test 74 "Ensure applications don't fail waiting for OST recovery"
2024 remote_dir_check_80() {
2026 local diridx=$($GETSTRIPE -M $remote_dir)
2027 [ $diridx -eq $MDTIDX ] || error "$diridx != $MDTIDX"
2029 createmany -o $remote_dir/f-%d 20 || error "creation failed"
2030 local fileidx=$($GETSTRIPE -M $remote_dir/f-1)
2031 [ $fileidx -eq $MDTIDX ] || error "$fileidx != $MDTIDX"
2037 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2038 ([ $FAILURE_MODE == "HARD" ] &&
2039 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2040 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2044 local remote_dir=$DIR/$tdir/remote_dir
2047 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2048 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2049 $LFS mkdir -i $MDTIDX $remote_dir &
2054 wait $CLIENT_PID || error "remote creation failed"
2056 remote_dir_check_80 || error "remote dir check failed"
2057 rm -rf $DIR/$tdir || error "rmdir failed"
2061 run_test 80a "DNE: create remote dir, drop update rep from MDT0, fail MDT0"
2064 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2065 ([ $FAILURE_MODE == "HARD" ] &&
2066 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2067 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2071 local remote_dir=$DIR/$tdir/remote_dir
2074 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2075 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2076 $LFS mkdir -i $MDTIDX $remote_dir &
2079 fail mds$((MDTIDX + 1))
2081 wait $CLIENT_PID || error "remote creation failed"
2083 remote_dir_check_80 || error "remote dir check failed"
2084 rm -rf $DIR/$tdir || error "rmdir failed"
2088 run_test 80b "DNE: create remote dir, drop update rep from MDT0, fail MDT1"
2091 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2092 ([ $FAILURE_MODE == "HARD" ] &&
2093 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2094 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2098 local remote_dir=$DIR/$tdir/remote_dir
2101 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2102 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2103 $LFS mkdir -i $MDTIDX $remote_dir &
2107 fail mds$((MDTIDX + 1))
2109 wait $CLIENT_PID || error "remote creation failed"
2111 remote_dir_check_80 || error "remote dir check failed"
2112 rm -rf $DIR/$tdir || error "rmdir failed"
2116 run_test 80c "DNE: create remote dir, drop update rep from MDT1, fail MDT[0,1]"
2119 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2121 local remote_dir=$DIR/$tdir/remote_dir
2124 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2125 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2126 $LFS mkdir -i $MDTIDX $remote_dir &
2129 # sleep 3 seconds to make sure MDTs are failed after
2130 # lfs mkdir -i has finished on all of MDTs.
2132 fail mds${MDTIDX},mds$((MDTIDX + 1))
2134 wait $CLIENT_PID || error "remote creation failed"
2136 remote_dir_check_80 || error "remote dir check failed"
2137 rm -rf $DIR/$tdir || error "rmdir failed"
2141 run_test 80d "DNE: create remote dir, drop update rep from MDT1, fail 2 MDTs"
2144 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2145 ([ $FAILURE_MODE == "HARD" ] &&
2146 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2147 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2151 local remote_dir=$DIR/$tdir/remote_dir
2154 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2155 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2156 $LFS mkdir -i $MDTIDX $remote_dir &
2159 # sleep 3 seconds to make sure MDTs are failed after
2160 # lfs mkdir -i has finished on all of MDTs.
2165 wait $CLIENT_PID || error "remote creation failed"
2167 remote_dir_check_80 || error "remote dir check failed"
2168 rm -rf $DIR/$tdir || error "rmdir failed"
2172 run_test 80e "DNE: create remote dir, drop MDT1 rep, fail MDT0"
2175 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2176 ([ $FAILURE_MODE == "HARD" ] &&
2177 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2178 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2181 local remote_dir=$DIR/$tdir/remote_dir
2184 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2185 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2186 $LFS mkdir -i $MDTIDX $remote_dir &
2189 fail mds$((MDTIDX + 1))
2191 wait $CLIENT_PID || error "remote creation failed"
2193 remote_dir_check_80 || error "remote dir check failed"
2194 rm -rf $DIR/$tdir || error "rmdir failed"
2198 run_test 80f "DNE: create remote dir, drop MDT1 rep, fail MDT1"
2201 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2202 ([ $FAILURE_MODE == "HARD" ] &&
2203 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2204 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2208 local remote_dir=$DIR/$tdir/remote_dir
2211 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2212 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2213 $LFS mkdir -i $MDTIDX $remote_dir &
2216 # sleep 3 seconds to make sure MDTs are failed after
2217 # lfs mkdir -i has finished on all of MDTs.
2221 fail mds$((MDTIDX + 1))
2223 wait $CLIENT_PID || error "remote creation failed"
2225 remote_dir_check_80 || error "remote dir check failed"
2226 rm -rf $DIR/$tdir || error "rmdir failed"
2230 run_test 80g "DNE: create remote dir, drop MDT1 rep, fail MDT0, then MDT1"
2233 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2235 local remote_dir=$DIR/$tdir/remote_dir
2238 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2239 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2240 $LFS mkdir -i $MDTIDX $remote_dir &
2243 # sleep 3 seconds to make sure MDTs are failed after
2244 # lfs mkdir -i has finished on all of MDTs.
2247 fail mds${MDTIDX},mds$((MDTIDX + 1))
2249 wait $CLIENT_PID || return 1
2251 remote_dir_check_80 || error "remote dir check failed"
2252 rm -rf $DIR/$tdir || error "rmdir failed"
2256 run_test 80h "DNE: create remote dir, drop MDT1 rep, fail 2 MDTs"
2259 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2260 ([ $FAILURE_MODE == "HARD" ] &&
2261 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2262 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2266 local remote_dir=$DIR/$tdir/remote_dir
2269 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2272 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2273 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2277 fail mds$((MDTIDX + 1))
2279 wait $CLIENT_PID || error "rm remote dir failed"
2281 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2283 rm -rf $DIR/$tdir || error "rmdir failed"
2287 run_test 81a "DNE: unlink remote dir, drop MDT0 update rep, fail MDT1"
2290 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2291 ([ $FAILURE_MODE == "HARD" ] &&
2292 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2293 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2296 local remote_dir=$DIR/$tdir/remote_dir
2299 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2301 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2302 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
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 81b "DNE: unlink remote dir, drop MDT0 update reply, fail MDT0"
2319 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2320 ([ $FAILURE_MODE == "HARD" ] &&
2321 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2322 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2326 local remote_dir=$DIR/$tdir/remote_dir
2329 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2331 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2332 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2337 fail mds$((MDTIDX + 1))
2339 wait $CLIENT_PID || error "rm remote dir failed"
2341 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2343 rm -rf $DIR/$tdir || error "rmdir failed"
2347 run_test 81c "DNE: unlink remote dir, drop MDT0 update reply, fail MDT0,MDT1"
2350 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2352 local remote_dir=$DIR/$tdir/remote_dir
2355 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2357 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2358 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2362 fail mds${MDTIDX},mds$((MDTIDX + 1))
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 81d "DNE: unlink remote dir, drop MDT0 update reply, fail 2 MDTs"
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
2391 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0
2395 wait $CLIENT_PID || error "rm remote dir failed"
2397 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2399 rm -rf $DIR/$tdir || error "rmdir failed"
2403 run_test 81e "DNE: unlink remote dir, drop MDT1 req reply, fail MDT0"
2406 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2407 ([ $FAILURE_MODE == "HARD" ] &&
2408 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2409 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2413 local remote_dir=$DIR/$tdir/remote_dir
2416 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2418 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2419 do_facet mds$((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 81f "DNE: unlink remote dir, drop MDT1 req reply, fail MDT1"
2436 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2437 ([ $FAILURE_MODE == "HARD" ] &&
2438 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2439 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2443 local remote_dir=$DIR/$tdir/remote_dir
2446 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2448 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2449 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2454 fail mds$((MDTIDX + 1))
2456 wait $CLIENT_PID || error "rm remote dir failed"
2458 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2460 rm -rf $DIR/$tdir || error "rmdir failed"
2464 run_test 81g "DNE: unlink remote dir, drop req reply, fail M0, then M1"
2467 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2469 local remote_dir=$DIR/$tdir/remote_dir
2472 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2474 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2475 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2479 fail mds${MDTIDX},mds$((MDTIDX + 1))
2481 wait $CLIENT_PID || error "rm remote dir failed"
2483 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2485 rm -rf $DIR/$tdir || error "rmdir failed"
2489 run_test 81h "DNE: unlink remote dir, drop request reply, fail 2 MDTs"
2493 createmany -o $DIR/$tdir/$tfile- 10 || return 1
2494 #define OBD_FAIL_MDS_FAIL_LOV_LOG_ADD 0x140
2495 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000140"
2496 unlinkmany $DIR/$tdir/$tfile- 10 || return 2
2498 run_test 83a "fail log_add during unlink recovery"
2502 createmany -o $DIR/$tdir/$tfile- 10 || return 1
2503 replay_barrier $SINGLEMDS
2504 unlinkmany $DIR/$tdir/$tfile- 10 || return 2
2505 #define OBD_FAIL_MDS_FAIL_LOV_LOG_ADD 0x140
2506 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000140"
2509 run_test 83b "fail log_add during unlink recovery"
2512 #define OBD_FAIL_MDS_OPEN_WAIT_CREATE 0x144
2513 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000144"
2514 createmany -o $DIR/$tfile- 1 &
2518 client_up || client_up || true # reconnect
2520 run_test 84a "stale open during export disconnect"
2522 test_85a() { #bug 16774
2523 lctl set_param -n ldlm.cancel_unused_locks_before_replay "1"
2525 for i in `seq 100`; do
2526 echo "tag-$i" > $DIR/$tfile-$i
2527 grep -q "tag-$i" $DIR/$tfile-$i || error "f2-$i"
2530 lov_id=`lctl dl | grep "clilov"`
2531 addr=`echo $lov_id | awk '{print $4}' | awk -F '-' '{print $3}'`
2532 count=`lctl get_param -n ldlm.namespaces.*MDT0000*$addr.lock_unused_count`
2533 echo "before recovery: unused locks count = $count"
2537 count2=`lctl get_param -n ldlm.namespaces.*MDT0000*$addr.lock_unused_count`
2538 echo "after recovery: unused locks count = $count2"
2540 if [ $count2 -ge $count ]; then
2541 error "unused locks are not canceled"
2544 run_test 85a "check the cancellation of unused locks during recovery(IBITS)"
2546 test_85b() { #bug 16774
2547 lctl set_param -n ldlm.cancel_unused_locks_before_replay "1"
2549 do_facet mgs $LCTL pool_new $FSNAME.$TESTNAME || return 1
2550 do_facet mgs $LCTL pool_add $FSNAME.$TESTNAME $FSNAME-OST0000 || return 2
2552 $SETSTRIPE -c 1 -p $FSNAME.$TESTNAME $DIR
2554 for i in `seq 100`; do
2555 dd if=/dev/urandom of=$DIR/$tfile-$i bs=4096 count=32 >/dev/null 2>&1
2558 cancel_lru_locks osc
2560 for i in `seq 100`; do
2561 dd if=$DIR/$tfile-$i of=/dev/null bs=4096 count=32 >/dev/null 2>&1
2564 lov_id=`lctl dl | grep "clilov"`
2565 addr=`echo $lov_id | awk '{print $4}' | awk -F '-' '{print $3}'`
2566 count=`lctl get_param -n ldlm.namespaces.*OST0000*$addr.lock_unused_count`
2567 echo "before recovery: unused locks count = $count"
2568 [ $count != 0 ] || return 3
2572 count2=`lctl get_param -n ldlm.namespaces.*OST0000*$addr.lock_unused_count`
2573 echo "after recovery: unused locks count = $count2"
2575 do_facet mgs $LCTL pool_remove $FSNAME.$TESTNAME $FSNAME-OST0000 || return 4
2576 do_facet mgs $LCTL pool_destroy $FSNAME.$TESTNAME || return 5
2578 if [ $count2 -ge $count ]; then
2579 error "unused locks are not canceled"
2582 run_test 85b "check the cancellation of unused locks during recovery(EXTENT)"
2585 local clients=${CLIENTS:-$HOSTNAME}
2587 zconf_umount_clients $clients $MOUNT
2588 do_facet $SINGLEMDS lctl set_param mdt.${FSNAME}-MDT*.exports.clear=0
2589 remount_facet $SINGLEMDS
2590 zconf_mount_clients $clients $MOUNT
2592 run_test 86 "umount server after clear nid_stats should not hit LBUG"
2595 do_facet ost1 "lctl set_param -n obdfilter.${ost1_svc}.sync_journal 0"
2598 $SETSTRIPE -i 0 -c 1 $DIR/$tfile
2599 dd if=/dev/urandom of=$DIR/$tfile bs=1024k count=8 || error "Cannot write"
2600 cksum=`md5sum $DIR/$tfile | awk '{print $1}'`
2601 cancel_lru_locks osc
2603 dd if=$DIR/$tfile of=/dev/null bs=1024k count=8 || error "Cannot read"
2604 cksum2=`md5sum $DIR/$tfile | awk '{print $1}'`
2605 if [ $cksum != $cksum2 ] ; then
2606 error "New checksum $cksum2 does not match original $cksum"
2609 run_test 87 "write replay"
2612 do_facet ost1 "lctl set_param -n obdfilter.${ost1_svc}.sync_journal 0"
2615 $SETSTRIPE -i 0 -c 1 $DIR/$tfile
2616 dd if=/dev/urandom of=$DIR/$tfile bs=1024k count=8 || error "Cannot write"
2617 sleep 1 # Give it a chance to flush dirty data
2618 echo TESTTEST | dd of=$DIR/$tfile bs=1 count=8 seek=64
2619 cksum=`md5sum $DIR/$tfile | awk '{print $1}'`
2620 cancel_lru_locks osc
2622 dd if=$DIR/$tfile of=/dev/null bs=1024k count=8 || error "Cannot read"
2623 cksum2=`md5sum $DIR/$tfile | awk '{print $1}'`
2624 if [ $cksum != $cksum2 ] ; then
2625 error "New checksum $cksum2 does not match original $cksum"
2628 run_test 87b "write replay with changed data (checksum resend)"
2630 test_88() { #bug 17485
2634 $SETSTRIPE -i 0 -c 1 $DIR/$tdir || error "$SETSTRIPE"
2637 replay_barrier $SINGLEMDS
2639 # exhaust precreations on ost1
2640 local OST=$(ostname_from_index 0)
2641 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $OST)
2642 local last_id=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_last_id)
2643 local next_id=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_next_id)
2644 echo "before test: last_id = $last_id, next_id = $next_id"
2646 echo "Creating to objid $last_id on ost $OST..."
2647 createmany -o $DIR/$tdir/f-%d $next_id $((last_id - next_id + 2))
2649 #create some files to use some uncommitted objids
2650 last_id=$(($last_id + 1))
2651 createmany -o $DIR/$tdir/f-%d $last_id 8
2653 last_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_last_id)
2654 next_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_next_id)
2655 echo "before recovery: last_id = $last_id2, next_id = $next_id2"
2657 # if test uses shutdown_facet && reboot_facet instead of facet_failover ()
2658 # it has to take care about the affected facets, bug20407
2659 local affected_mds1=$(affected_facets mds1)
2660 local affected_ost1=$(affected_facets ost1)
2662 shutdown_facet $SINGLEMDS
2665 reboot_facet $SINGLEMDS
2666 change_active $affected_mds1
2667 wait_for_facet $affected_mds1
2668 mount_facets $affected_mds1 || error "Restart of mds failed"
2671 change_active $affected_ost1
2672 wait_for_facet $affected_ost1
2673 mount_facets $affected_ost1 || error "Restart of ost1 failed"
2677 last_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_last_id)
2678 next_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_next_id)
2679 echo "after recovery: last_id = $last_id2, next_id = $next_id2"
2681 # create new files, which should use new objids, and ensure the orphan
2682 # cleanup phase for ost1 is completed at the same time
2683 for i in `seq 8`; do
2684 file_id=$(($last_id + 10 + $i))
2685 dd if=/dev/urandom of=$DIR/$tdir/f-$file_id bs=4096 count=128
2688 # if the objids were not recreated, then "ls" will failed for -ENOENT
2689 ls -l $DIR/$tdir/* || error "can't get the status of precreated files"
2692 # write into previously created files
2693 for i in `seq 8`; do
2694 file_id=$(($last_id + $i))
2695 dd if=/dev/urandom of=$DIR/$tdir/f-$file_id bs=4096 count=128
2696 cp -f $DIR/$tdir/f-$file_id $TMP/$tdir/
2699 # compare the content
2700 for i in `seq 8`; do
2701 file_id=$(($last_id + $i))
2702 cmp $TMP/$tdir/f-$file_id $DIR/$tdir/f-$file_id || error "the content" \
2703 "of file is modified!"
2708 run_test 88 "MDS should not assign same objid to different files "
2711 cancel_lru_locks osc
2713 rm -f $DIR/$tdir/$tfile
2715 wait_delete_completed
2716 BLOCKS1=$(df -P $MOUNT | tail -n 1 | awk '{ print $3 }')
2717 $SETSTRIPE -i 0 -c 1 $DIR/$tdir/$tfile
2718 dd if=/dev/zero bs=1M count=10 of=$DIR/$tdir/$tfile
2721 facet_failover $SINGLEMDS
2722 rm $DIR/$tdir/$tfile
2725 zconf_mount $(hostname) $MOUNT
2726 client_up || return 1
2728 wait_delete_completed
2729 BLOCKS2=$(df -P $MOUNT | tail -n 1 | awk '{ print $3 }')
2730 [ $((BLOCKS2 - BLOCKS1)) -le 4 ] || \
2731 error $((BLOCKS2 - BLOCKS1)) blocks leaked
2734 run_test 89 "no disk space leak on late ost connection"
2740 change_active $facet
2741 wait_for_facet $facet
2742 mount_facet $facet || error "Restart of $facet failed"
2746 test_90() { # bug 19494
2747 local dir=$DIR/$tdir
2748 local ostfail=$(get_random_entry $(get_facets OST))
2750 if [[ $FAILURE_MODE = HARD ]]; then
2751 local affected=$(affected_facets $ostfail);
2752 if [[ "$affected" != $ostfail ]]; then
2753 skip not functional with FAILURE_MODE=$FAILURE_MODE, affected: $affected
2760 echo "Create the files"
2762 # file "f${index}" striped over 1 OST
2763 # file "all" striped over all OSTs
2765 $SETSTRIPE -c $OSTCOUNT $dir/all ||
2766 error "setstripe failed to create $dir/all"
2768 for (( i=0; i<$OSTCOUNT; i++ )); do
2770 $SETSTRIPE -i $i -c 1 $f || error "$SETSTRIPE failed to create $f"
2772 # confirm setstripe actually created the stripe on the requested OST
2773 local uuid=$(ostuuid_from_index $i)
2774 for file in f$i all; do
2775 if [[ $dir/$file != $($LFS find --obd $uuid --name $file $dir) ]]; then
2776 $GETSTRIPE $dir/$file
2777 error wrong stripe: $file, uuid: $uuid
2782 # Before failing an OST, get its obd name and index
2783 local varsvc=${ostfail}_svc
2784 local obd=$(do_facet $ostfail lctl get_param -n obdfilter.${!varsvc}.uuid)
2785 local index=$(($(facet_number $ostfail) - 1))
2787 echo "Fail $ostfail $obd, display the list of affected files"
2788 shutdown_facet $ostfail || return 2
2790 trap "cleanup_90 $ostfail" EXIT INT
2791 echo "General Query: lfs find $dir"
2792 local list=$($LFS find $dir)
2794 for (( i=0; i<$OSTCOUNT; i++ )); do
2795 list_member "$list" $dir/f$i || error_noexit "lfs find $dir: no file f$i"
2797 list_member "$list" $dir/all || error_noexit "lfs find $dir: no file all"
2799 # focus on the missing OST,
2800 # we expect to see only two files affected: "f$(index)" and "all"
2802 echo "Querying files on shutdown $ostfail: lfs find --obd $obd"
2803 list=$($LFS find --obd $obd $dir)
2805 for file in all f$index; do
2806 list_member "$list" $dir/$file ||
2807 error_noexit "lfs find does not report the affected $obd for $file"
2810 [[ $(echo $list | wc -w) -eq 2 ]] ||
2811 error_noexit "lfs find reports the wrong list of affected files ${#list[@]}"
2813 echo "Check getstripe: $GETSTRIPE -r --obd $obd"
2814 list=$($GETSTRIPE -r --obd $obd $dir)
2816 for file in all f$index; do
2817 echo "$list" | grep $dir/$file ||
2818 error_noexit "lfs getsripe does not report the affected $obd for $file"
2823 run_test 90 "lfs find identifies the missing striped file segments"
2826 check_and_cleanup_lustre