6 # This test needs to be run on the client
9 LUSTRE=${LUSTRE:-`dirname $0`/..}
10 . $LUSTRE/tests/test-framework.sh
14 . ${CONFIG:=$LUSTRE/tests/cfg/local.sh}
22 add_mds mds --dev $MDSDEV --size $MDSSIZE
23 if [ ! -z "$mdsfailover_HOST" ]; then
24 add_mdsfailover mds --dev $MDSDEV --size $MDSSIZE
27 add_lov lov1 mds --stripe_sz $STRIPE_BYTES \
28 --stripe_cnt $STRIPES_PER_OBJ --stripe_pattern 0
29 add_ost ost --lov lov1 --dev $OSTDEV --size $OSTSIZE
30 add_ost ost2 --lov lov1 --dev ${OSTDEV}-2 --size $OSTSIZE
31 add_client client mds --lov lov1 --path $MOUNT
37 # make sure we are using the primary MDS, so the config log will
38 # be able to clean up properly.
39 activemds=`facet_active mds`
40 if [ $activemds != "mds" ]; then
43 zconf_umount `hostname` $MOUNT
44 stop mds ${FORCE} $MDSLCONFARGS
46 stop ost ${FORCE} --dump $TMP/replay-single-`hostname`.log
49 if [ "$ONLY" == "cleanup" ]; then
50 sysctl -w portals.debug=0 || true
55 SETUP=${SETUP:-"setup"}
56 CLEANUP=${CLEANUP:-"cleanup"}
61 start ost --reformat $OSTLCONFARGS
62 start ost2 --reformat $OSTLCONFARGS
63 [ "$DAEMONFILE" ] && $LCTL debug_daemon start $DAEMONFILE $DAEMONSIZE
64 start mds $MDSLCONFARGS --reformat
65 grep " $MOUNT " /proc/mounts || zconf_mount `hostname` $MOUNT
70 if [ "$ONLY" == "setup" ]; then
80 run_test 0 "empty replay"
83 # this test attempts to trigger a race in the precreation code,
84 # and must run before any other objects are created on the filesystem
86 createmany -o $DIR/$tfile 20 || return 1
87 unlinkmany $DIR/$tfile 20 || return 2
89 run_test 0b "ensure object created after recover exists. (3284)"
95 $CHECKSTAT -t file $DIR/$tfile || return 1
98 run_test 1 "simple create"
104 $CHECKSTAT -t file $DIR/$tfile || return 1
110 ./mcreate $DIR/$tfile
114 $CHECKSTAT -t file $DIR/$tfile || return 1
122 o_directory $DIR/$tfile
124 $CHECKSTAT -t file $DIR/$tfile || return 2
127 run_test 3a "replay failed open(O_DIRECTORY)"
131 #define OBD_FAIL_MDS_OPEN_PACK | OBD_FAIL_ONCE
132 do_facet mds "sysctl -w lustre.fail_loc=0x80000114"
134 do_facet mds "sysctl -w lustre.fail_loc=0"
136 $CHECKSTAT -t file $DIR/$tfile && return 2
139 run_test 3b "replay failed open -ENOMEM"
143 #define OBD_FAIL_MDS_ALLOC_OBDO | OBD_FAIL_ONCE
144 do_facet mds "sysctl -w lustre.fail_loc=0x80000128"
146 do_facet mds "sysctl -w lustre.fail_loc=0"
149 $CHECKSTAT -t file $DIR/$tfile && return 2
152 run_test 3c "replay failed open -ENOMEM"
156 for i in `seq 10`; do
157 echo "tag-$i" > $DIR/$tfile-$i
160 for i in `seq 10`; do
161 grep -q "tag-$i" $DIR/$tfile-$i || error "$tfile-$i"
164 run_test 4 "|x| 10 open(O_CREAT)s"
170 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
172 run_test 4b "|x| rm 10 files"
174 # The idea is to get past the first block of precreated files on both
175 # osts, and then replay.
178 for i in `seq 220`; do
179 echo "tag-$i" > $DIR/$tfile-$i
182 for i in `seq 220`; do
183 grep -q "tag-$i" $DIR/$tfile-$i || error "f1c-$i"
187 # waiting for commitment of removal
189 run_test 5 "|x| 220 open(O_CREAT)"
195 mcreate $DIR/$tdir/$tfile
197 $CHECKSTAT -t dir $DIR/$tdir || return 1
198 $CHECKSTAT -t file $DIR/$tdir/$tfile || return 2
200 # waiting for log process thread
202 run_test 6 "mkdir + contained create"
208 $CHECKSTAT -t dir $DIR/$tdir && return 1 || true
210 run_test 6b "|X| rmdir"
215 mcreate $DIR/$tdir/$tfile
217 $CHECKSTAT -t dir $DIR/$tdir || return 1
218 $CHECKSTAT -t file $DIR/$tdir/$tfile || return 2
221 run_test 7 "mkdir |X| contained create"
225 multiop $DIR/$tfile mo_c &
230 $CHECKSTAT -t file $DIR/$tfile || return 1
231 kill -USR1 $MULTIPID || return 2
232 wait $MULTIPID || return 3
235 run_test 8 "creat open |X| close"
240 local old_inum=`ls -i $DIR/$tfile | awk '{print $1}'`
242 local new_inum=`ls -i $DIR/$tfile | awk '{print $1}'`
244 echo " old_inum == $old_inum, new_inum == $new_inum"
245 if [ $old_inum -eq $new_inum ] ;
247 echo " old_inum and new_inum match"
249 echo "!!!! old_inum and new_inum NOT match"
254 run_test 9 "|X| create (same inum/gen)"
259 mv $DIR/$tfile $DIR/$tfile-2
262 $CHECKSTAT $DIR/$tfile && return 1
263 $CHECKSTAT $DIR/$tfile-2 ||return 2
267 run_test 10 "create |X| rename unlink"
271 echo "old" > $DIR/$tfile
272 mv $DIR/$tfile $DIR/$tfile-2
274 echo "new" > $DIR/$tfile
276 grep old $DIR/$tfile-2
278 grep new $DIR/$tfile || return 1
279 grep old $DIR/$tfile-2 || return 2
281 run_test 11 "create open write rename |X| create-old-name read"
285 multiop $DIR/$tfile o_tSc &
287 # give multiop a chance to open
292 wait $pid || return 1
295 [ -e $DIR/$tfile ] && return 2
298 run_test 12 "open, unlink |X| close"
301 # 1777 - replay open after committed chmod that would make
302 # a regular open a failure
305 multiop $DIR/$tfile O_wc &
307 # give multiop a chance to open
310 $CHECKSTAT -p 0 $DIR/$tfile
314 wait $pid || return 1
316 $CHECKSTAT -s 1 -p 0 $DIR/$tfile || return 2
319 run_test 13 "open chmod 0 |x| write close"
322 multiop $DIR/$tfile O_tSc &
324 # give multiop a chance to open
328 kill -USR1 $pid || return 1
329 wait $pid || return 2
332 [ -e $DIR/$tfile ] && return 3
335 run_test 14 "open(O_CREAT), unlink |X| close"
338 multiop $DIR/$tfile O_tSc &
340 # give multiop a chance to open
344 touch $DIR/g11 || return 1
346 wait $pid || return 2
349 [ -e $DIR/$tfile ] && return 3
350 touch $DIR/h11 || return 4
353 run_test 15 "open(O_CREAT), unlink |X| touch new, close"
360 mcreate $DIR/$tfile-2
362 [ -e $DIR/$tfile ] && return 1
363 [ -e $DIR/$tfile-2 ] || return 2
364 munlink $DIR/$tfile-2 || return 3
366 run_test 16 "|X| open(O_CREAT), unlink, touch new, unlink new"
370 multiop $DIR/$tfile O_c &
372 # give multiop a chance to open
375 kill -USR1 $pid || return 1
376 wait $pid || return 2
377 $CHECKSTAT -t file $DIR/$tfile || return 3
380 run_test 17 "|X| open(O_CREAT), |replay| close"
384 multiop $DIR/$tfile O_tSc &
386 # give multiop a chance to open
389 touch $DIR/$tfile-2 || return 1
390 echo "pid: $pid will close"
392 wait $pid || return 2
395 [ -e $DIR/$tfile ] && return 3
396 [ -e $DIR/$tfile-2 ] || return 4
397 # this touch frequently fails
398 touch $DIR/$tfile-3 || return 5
399 munlink $DIR/$tfile-2 || return 6
400 munlink $DIR/$tfile-3 || return 7
403 run_test 18 "|X| open(O_CREAT), unlink, touch new, close, touch, unlink"
405 # bug 1855 (a simpler form of test_11 above)
409 echo "old" > $DIR/$tfile
410 mv $DIR/$tfile $DIR/$tfile-2
411 grep old $DIR/$tfile-2
413 grep old $DIR/$tfile-2 || return 2
415 run_test 19 "|X| mcreate, open, write, rename "
419 multiop $DIR/$tfile O_tSc &
421 # give multiop a chance to open
427 wait $pid || return 1
428 [ -e $DIR/$tfile ] && return 2
431 run_test 20 "|X| open(O_CREAT), unlink, replay, close (test mds_cleanup_orphans)"
435 multiop $DIR/$tfile O_tSc &
437 # give multiop a chance to open
440 touch $DIR/g11 || return 1
444 wait $pid || return 2
445 [ -e $DIR/$tfile ] && return 3
446 touch $DIR/h11 || return 4
449 run_test 21 "|X| open(O_CREAT), unlink touch new, replay, close (test mds_cleanup_orphans)"
452 multiop $DIR/$tfile O_tSc &
454 # give multiop a chance to open
462 wait $pid || return 1
463 [ -e $DIR/$tfile ] && return 2
466 run_test 22 "open(O_CREAT), |X| unlink, replay, close (test mds_cleanup_orphans)"
469 multiop $DIR/$tfile O_tSc &
471 # give multiop a chance to open
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 23 "open(O_CREAT), |X| unlink touch new, replay, close (test mds_cleanup_orphans)"
488 multiop $DIR/$tfile O_tSc &
490 # give multiop a chance to open
497 wait $pid || return 1
498 [ -e $DIR/$tfile ] && return 2
501 run_test 24 "open(O_CREAT), replay, unlink, close (test mds_cleanup_orphans)"
504 multiop $DIR/$tfile O_tSc &
506 # give multiop a chance to open
513 wait $pid || return 1
514 [ -e $DIR/$tfile ] && return 2
517 run_test 25 "open(O_CREAT), unlink, replay, close (test mds_cleanup_orphans)"
521 multiop $DIR/$tfile-1 O_tSc &
523 multiop $DIR/$tfile-2 O_tSc &
525 # give multiop a chance to open
530 wait $pid2 || return 1
534 wait $pid1 || return 2
535 [ -e $DIR/$tfile-1 ] && return 3
536 [ -e $DIR/$tfile-2 ] && return 4
539 run_test 26 "|X| open(O_CREAT), unlink two, close one, replay, close one (test mds_cleanup_orphans)"
543 multiop $DIR/$tfile-1 O_tSc &
545 multiop $DIR/$tfile-2 O_tSc &
547 # give multiop a chance to open
554 wait $pid1 || return 1
556 wait $pid2 || return 2
557 [ -e $DIR/$tfile-1 ] && return 3
558 [ -e $DIR/$tfile-2 ] && return 4
561 run_test 27 "|X| open(O_CREAT), unlink two, replay, close two (test mds_cleanup_orphans)"
564 multiop $DIR/$tfile-1 O_tSc &
566 multiop $DIR/$tfile-2 O_tSc &
568 # give multiop a chance to open
574 wait $pid2 || return 1
578 wait $pid1 || return 2
579 [ -e $DIR/$tfile-1 ] && return 3
580 [ -e $DIR/$tfile-2 ] && return 4
583 run_test 28 "open(O_CREAT), |X| unlink two, close one, replay, close one (test mds_cleanup_orphans)"
586 multiop $DIR/$tfile-1 O_tSc &
588 multiop $DIR/$tfile-2 O_tSc &
590 # give multiop a chance to open
598 wait $pid1 || return 1
600 wait $pid2 || return 2
601 [ -e $DIR/$tfile-1 ] && return 3
602 [ -e $DIR/$tfile-2 ] && return 4
605 run_test 29 "open(O_CREAT), |X| unlink two, replay, close two (test mds_cleanup_orphans)"
608 multiop $DIR/$tfile-1 O_tSc &
610 multiop $DIR/$tfile-2 O_tSc &
612 # give multiop a chance to open
620 wait $pid1 || return 1
622 wait $pid2 || return 2
623 [ -e $DIR/$tfile-1 ] && return 3
624 [ -e $DIR/$tfile-2 ] && return 4
627 run_test 30 "open(O_CREAT) two, unlink two, replay, close two (test mds_cleanup_orphans)"
630 multiop $DIR/$tfile-1 O_tSc &
632 multiop $DIR/$tfile-2 O_tSc &
634 # give multiop a chance to open
642 wait $pid1 || return 1
644 wait $pid2 || return 2
645 [ -e $DIR/$tfile-1 ] && return 3
646 [ -e $DIR/$tfile-2 ] && return 4
649 run_test 31 "open(O_CREAT) two, unlink one, |X| unlink one, close two (test mds_cleanup_orphans)"
651 # tests for bug 2104; completion without crashing is success. The close is
652 # stale, but we always return 0 for close, so the app never sees it.
654 multiop $DIR/$tfile O_c &
656 multiop $DIR/$tfile O_c &
658 # give multiop a chance to open
661 df $MOUNT || sleep 1 && df $MOUNT || return 1
667 run_test 32 "close() notices client eviction; close() after client eviction"
669 # Abort recovery before client complete
674 # this file should be gone, because the replay was aborted
675 $CHECKSTAT -t file $DIR/$tfile && return 1
678 run_test 33 "abort recovery before client does replay"
681 multiop $DIR/$tfile O_c &
683 # give multiop a chance to open
690 [ -e $DIR/$tfile ] && return 1
694 run_test 34 "abort recovery before client does replay (test mds_cleanup_orphans)"
696 # bug 2278 - generate one orphan on OST, then destroy it during recovery from llog
700 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
701 do_facet mds "sysctl -w lustre.fail_loc=0x80000119"
706 # give a chance to remove from MDS
708 $CHECKSTAT -t file $DIR/$tfile && return 1 || true
710 run_test 35 "test recovery from llog for unlink op"
712 # b=2432 resent cancel after replay uses wrong cookie,
713 # so don't resend cancels
717 checkstat $DIR/$tfile
720 if dmesg | grep "unknown lock cookie"; then
721 echo "cancel after replay failed"
725 run_test 36 "don't resend cancel"
728 # directory orphans can't be unlinked from PENDING directory
730 rmdir $DIR/$tfile 2>/dev/null
731 multiop $DIR/$tfile dD_c &
733 # give multiop a chance to open
738 # clear the dmesg buffer so we only see errors from this recovery
742 dmesg | grep "mds_unlink_orphan.*error .* unlinking orphan" && return 1
746 run_test 37 "abort recovery before client does replay (test mds_cleanup_orphans for directories)"
749 createmany -o $DIR/$tfile-%d 800
750 unlinkmany $DIR/$tfile-%d 0 400
753 unlinkmany $DIR/$tfile-%d 400 400
755 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
757 run_test 38 "test recovery from unlink llog (test llog_gen_rec) "
760 createmany -o $DIR/$tfile-%d 800
762 unlinkmany $DIR/$tfile-%d 0 400
764 unlinkmany $DIR/$tfile-%d 400 400
766 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
768 run_test 39 "test recovery from unlink llog (test llog_gen_rec) "
771 cat /proc/fs/lustre/osc/*/stats |
772 awk -vwrites=0 '/ost_write/ { writes += $2 } END { print writes; }'
777 $LCTL mark multiop $MOUNT/$tfile OS_c
778 multiop $MOUNT/$tfile OS_c &
780 writeme -s $MOUNT/${tfile}-2 &
784 #define OBD_FAIL_MDS_CONNECT_NET 0x117
785 do_facet mds "sysctl -w lustre.fail_loc=0x80000117"
787 stat1=`count_ost_writes`
789 stat2=`count_ost_writes`
790 echo "$stat1, $stat2"
791 if [ $stat1 -lt $stat2 ]; then
792 echo "writes continuing during recovery"
795 echo "writes not continuing during recovery, bug 2477"
798 echo "waiting for writeme $WRITE_PID"
802 echo "waiting for multiop $PID"
803 wait $PID || return 2
804 do_facet client munlink $MOUNT/$tfile || return 3
805 do_facet client munlink $MOUNT/${tfile}-2 || return 3
808 run_test 40 "cause recovery in ptlrpc, ensure IO continues"
812 # make sure that a read to one osc doesn't try to double-unlock its page just
813 # because another osc is invalid. trigger_group_io used to mistakenly return
814 # an error if any oscs were invalid even after having successfully put rpcs
815 # on valid oscs. This was fatal if the caller was ll_readpage who unlocked
816 # the page, guarnateeing that the unlock from the RPC completion would
817 # assert on trying to unlock the unlocked page.
819 local f=$MOUNT/$tfile
820 # make sure the start of the file is ost1
821 lfs setstripe $f $((128 * 1024)) 0 0
822 do_facet client dd if=/dev/zero of=$f bs=4k count=1 || return 3
824 # fail ost2 and read from ost1
825 local osc2_dev=`$LCTL device_list | \
826 awk '(/ost2.*client_facet/){print $4}' `
827 $LCTL --device %$osc2_dev deactivate
828 do_facet client dd if=$f of=/dev/null bs=4k count=1 || return 3
829 $LCTL --device %$osc2_dev activate
832 run_test 41 "read from a valid osc while other oscs are invalid"
834 # test MDS recovery after ost failure
836 blocks=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
837 createmany -o $DIR/$tfile-%d 800
839 unlinkmany $DIR/$tfile-%d 0 400
842 # osc is evicted, fs is smaller
843 blocks_after=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
844 [ $blocks_after -lt $blocks ] || return 1
845 echo wait for MDS to timeout and recover
846 sleep $((TIMEOUT * 2))
847 unlinkmany $DIR/$tfile-%d 400 400
848 $CHECKSTAT -t file $DIR/$tfile-* && return 2 || true
850 run_test 42 "recovery after ost failure"
853 # timeout in MDS/OST recovery RPC will LBUG MDS
857 # OBD_FAIL_OST_CREATE_NET 0x204
858 do_facet ost "sysctl -w lustre.fail_loc=0x80000204"
861 do_facet ost "sysctl -w lustre.fail_loc=0"
865 run_test 43 "mds osc import failure during recovery; don't LBUG"
868 mdcdev=`awk '/mds_svc_MNT/ {print $1}' < /proc/fs/lustre/devices`
869 do_facet mds "sysctl -w lustre.fail_loc=0x80000701"
870 $LCTL --device $mdcdev recover
872 do_facet mds "sysctl -w lustre.fail_loc=0"
875 run_test 44 "race in target handle connect"
877 # Handle failed close
879 mdcdev=`awk '/mds_svc_MNT/ {print $1}' < /proc/fs/lustre/devices`
880 $LCTL --device $mdcdev recover
882 multiop $DIR/$tfile O_c &
886 # This will cause the CLOSE to fail before even
887 # allocating a reply buffer
888 $LCTL --device $mdcdev deactivate
892 wait $pid || return 1
894 $LCTL --device $mdcdev activate
897 $CHECKSTAT -t file $DIR/$tfile || return 2
900 run_test 45 "Handle failed close"
904 drop_reply "touch $DIR/$tfile"
906 # ironically, the previous test, 45, will cause a real forced close,
907 # so just look for one for this test
908 dmesg | grep -i "force closing client file handle for $tfile" && return 1
911 run_test 46 "Don't leak file handle after open resend (3325)"
916 # create some files to make sure precreate has been done on all
917 # OSTs. (just in case this test is run independently)
918 createmany -o $DIR/$tfile 20 || return 1
920 # OBD_FAIL_OST_CREATE_NET 0x204
922 do_facet ost "sysctl -w lustre.fail_loc=0x80000204"
923 df $MOUNT || return 2
925 # let the MDS discover the OST failure, attempt to recover, fail
927 sleep $((3 * TIMEOUT))
929 # Without 2824, this createmany would hang
930 createmany -o $DIR/$tfile 20 || return 3
931 unlinkmany $DIR/$tfile 20 || return 4
933 do_facet ost "sysctl -w lustre.fail_loc=0"
936 run_test 47 "MDS->OSC failure during precreate cleanup (2824)"
941 createmany -o $DIR/$tfile 20 || return 1
942 # OBD_FAIL_OST_EROFS 0x216
944 do_facet ost "sysctl -w lustre.fail_loc=0x80000216"
945 df $MOUNT || return 2
947 createmany -o $DIR/$tfile 20 20 || return 2
948 unlinkmany $DIR/$tfile 40 || return 3
950 do_facet ost "sysctl -w lustre.fail_loc=0"
953 run_test 48 "MDS->OSC failure during precreate cleanup (2824)"
956 local osc_dev=`$LCTL device_list | \
957 awk '(/ost_svc_mds_svc/){print $4}' `
958 $LCTL --device %$osc_dev recover && $LCTL --device %$osc_dev recover
959 # give the mds_lov_sync threads a chance to run
962 run_test 50 "Double OSC recovery, don't LASSERT (3812)"
964 # b3764 timed out lock replay
969 multiop $DIR/$tfile s
971 do_facet mds "sysctl -w lustre.fail_loc=0x8000030c"
973 do_facet mds "sysctl -w lustre.fail_loc=0x0"
975 $CHECKSTAT -t file $DIR/$tfile-* && return 3 || true
977 run_test 52 "time out lock replay (3764)"
979 #b_cray 53 "|X| open request and close reply while two MDC requests in flight"
980 #b_cray 54 "|X| open request and close reply while two MDC requests in flight"
982 #b3761 ASSERTION(hash != 0) failed
984 # OBD_FAIL_MDS_OPEN_CREATE | OBD_FAIL_ONCE
985 do_facet mds "sysctl -w lustre.fail_loc=0x8000012b"
987 # give touch a chance to run
989 do_facet mds "sysctl -w lustre.fail_loc=0x0"
993 run_test 55 "let MDS_CHECK_RESENT return the original return code instead of 0"
995 #b3440 ASSERTION(rec->ur_fid2->id) failed
997 ln -s foo $DIR/$tfile
999 #drop_reply "cat $DIR/$tfile"
1003 run_test 56 "don't replay a symlink open request (3440)"
1005 equals_msg test complete, cleaning up