On Mon, Aug 26, 2019 at 02:20:45PM +0800, Qu Wenruo wrote:
> We have generic dm-logwrites with fsstress test case (generic/482), but
> it doesn't cover fs specific operations like btrfs snapshot creation and
> deletion.
>
> Furthermore, that test is not heavy enough to bump btrfs tree height by
> its short runtime.
>
> And finally, btrfs check doesn't consider dirty log as an error, unlike
> ext*/xfs, that's to say we don't need to mount the fs to replay the log,
> but just run btrfs check on the fs is enough.
>
> So introduce a similar test case but for btrfs only.
>
> The test case will stress btrfs by:
> - Use small nodesize to bump tree height
> - Create a base tree which is already high enough
> - Trim tree blocks to find possible trim bugs
> - Call snapshot creation and deletion along with fsstress
>
> To utilize replay-log --check and --fsck command, we fix one bug in
> replay-log first:
> - Return 1 when fsck failed
> Original when fsck failed, run_fsck() returns -1, but to make
> replay_log prog to return 1, we need to return a minus value, so
> fix it by setting @ret to -EUCLEAN when run_fsck() failed, so that
> we can detect the fsck failure by simply checking the return value
> of replay-log.
Sorry, I didn't quite get this. run_fsck() already returns a negative
value (-1) on fsck failure (thus @ret is -1 in this case), and
replay_log exits with 1 if @ret < 0. All seem fine to me, setting
-EUCLEAN doesn't seem necessary to me. Did I miss anything?
Anyway, I think this bugfix could be in a separate patch.
>
> Also it includes certain workaround for btrfs:
> - Use no-holes feature
> To avoid missing hole file extents.
> Although that behavior doesn't follow the on-disk format spec, it
> doesn't cause data loss. And will follow the new on-disk format spec
> of no-holes feature, so it's better to workaround it.
>
> And an optimization for btrfs only:
> - Use replay-log --fsck/--check command
> Since dm-log-writes records bios sequentially, there is no way to
> locate certain entry unless we iterate all entries.
> This is becoming a big performance penalty if we replay certain a
> range, check the fs, then re-execute replay-log to replay another
> range.
>
> We need to records the previous entry location, or we need to
> re-iterate all previous entries.
>
> Thankfully, replay-log has already address it by providing --fsck and
> --check command, thus we don't need to break replay-log command.
>
> Please note, for fast storage (e.g. fast NVME or unsafe cache mode),
> it's recommended to use log devices larger than 15G, or we can't record
> the full log of just 30s fsstress run.
>
> Signed-off-by: Qu Wenruo <wqu@xxxxxxxx>
> ---
> For the log devices size problem, I have submitted dm-logwrites bio flag
> filter support, to filter out data bios.
>
> But that is not yet merged into kernel, thus we need a large log device
> for short run.
>
> For reference, if using unsafe cache mode for all test devices, on a
> system with 32G dual-channel DDR4 3200 RAM, 5G log device will be
> filled up in less than 15 seconds.
> So to ensure dm-log-writes covers all the operations, one needs at least
> 15G log device, and even more if using RAM with more channels.
>
> Changelog:
> v2
> - Better expression/words for comment
> - Add requirement for no-holes features
> - Use xattr to bump up tree height
> So no need for max_inline mount option
> - Coding style fixes for function definition
> - Add -f for rm to avoid user alias setting
> - Add new workload (update time stamp and create new files) for snapshot
> workload
> - Remove an unnecessary sync call
> - Get rid of wrong 2>&1 redirection
> - Add to group "snapshot" and "stress"
>
> v3:
> - Add '_require_attrs' and source common/attr
> - Introduce '_require_fsck_not_report_dirty_logs_as_error'
> - Add comment for the replay-log code fix
> - Wait after killing all background fsstress
> - Use $BLKDISCARD_PROG instead of plain 'blkdiscard'
> - Add trap for snapshot and delete workload
> ---
> common/config | 1 +
> common/dmlogwrites | 44 ++++++++++
> src/log-writes/replay-log.c | 2 +
> tests/btrfs/192 | 156 ++++++++++++++++++++++++++++++++++++
> tests/btrfs/192.out | 2 +
> tests/btrfs/group | 1 +
> 6 files changed, 206 insertions(+)
> create mode 100755 tests/btrfs/192
> create mode 100644 tests/btrfs/192.out
>
> diff --git a/common/config b/common/config
> index bd64be62..4c86a492 100644
> --- a/common/config
> +++ b/common/config
> @@ -183,6 +183,7 @@ export LOGGER_PROG="$(type -P logger)"
> export DBENCH_PROG="$(type -P dbench)"
> export DMSETUP_PROG="$(type -P dmsetup)"
> export WIPEFS_PROG="$(type -P wipefs)"
> +export BLKDISCARD_PROG="$(type -P blkdiscard)"
> export DUMP_PROG="$(type -P dump)"
> export RESTORE_PROG="$(type -P restore)"
> export LVM_PROG="$(type -P lvm)"
> diff --git a/common/dmlogwrites b/common/dmlogwrites
> index ae2cbc6a..474ec570 100644
> --- a/common/dmlogwrites
> +++ b/common/dmlogwrites
> @@ -175,3 +175,47 @@ _log_writes_replay_log_range()
> >> $seqres.full 2>&1
> [ $? -ne 0 ] && _fail "replay failed"
> }
> +
> +# Require fsck not to report dirty logs as error
> +#
> +# This is a special requirement to use _log_writes_fast_replay_check
> +# The reasons are:
> +# - To avoid unnecessary seek when there are a lot of entries
> +# replay-log doesn't have a tree-like structure to do fast index,
> +# thus it iterate all entries one by one, this can be very slow
> +# - No way to revert the log replay for next check
> +# A lot of fsck will replay the log, which will pollute the replay device
> +# for next entry
> +_require_fsck_not_report_dirty_logs_as_error()
> +{
> + if [ $FSTYP != "btrfs" ]; then
> + _notrun "fsck of $FSTYP reports dirty jounal/log as error, skipping test"
> + fi
> +}
The rule name seems ugly :)
> +
> +# Replay and check each fua/flush (specified by $2) point.
> +#
> +# Since dm-log-writes records bio sequentially, even just replaying a range
> +# still needs to iterate all records before the end point.
> +# When number of records grows, it will be unacceptably slow, thus we need
> +# to use relay-log itself to trigger fsck, avoid unnecessary seek.
> +_log_writes_fast_replay_check()
> +{
> + local check_point=$1
> + local blkdev=$2
> + local fsck_command
> +
> + _require_fsck_not_report_dirty_logs_as_error
> +
> + [ -z "$check_point" -o -z "$blkdev" ] && _fail \
> + "check_point and blkdev must be specified for _log_writes_fast_replay_check"
> + case $FSTYP in
> + btrfs)
> + fsck_command="$BTRFS_UTIL_PROG check $blkdev"
> + ;;
> + esac
> + $here/src/log-writes/replay-log --log $LOGWRITES_DEV \
> + --replay $blkdev --check $check_point --fsck "$fsck_command" \
> + 2>&1 | tail -n 128 >> $seqres.full
> + [ $? -ne 0 ] && _fail "fsck failed during replay"
> +}
And I think we could make _log_writes_fast_replay_check(), which seems
only useful to btrfs, a local function in the test, so we avoid all
these $FSTYP == btrfs checks.
Thanks,
Eryu
> diff --git a/src/log-writes/replay-log.c b/src/log-writes/replay-log.c
> index 829b18e2..1e1cd524 100644
> --- a/src/log-writes/replay-log.c
> +++ b/src/log-writes/replay-log.c
> @@ -1,5 +1,6 @@
> // SPDX-License-Identifier: GPL-2.0
> #include <stdio.h>
> +#include <errno.h>
> #include <unistd.h>
> #include <getopt.h>
> #include <stdlib.h>
> @@ -375,6 +376,7 @@ int main(int argc, char **argv)
> fprintf(stderr, "Fsck errored out on entry "
> "%llu\n",
> (unsigned long long)log->cur_entry - 1);
> + ret = -EUCLEAN;
> break;
> }
> }
> diff --git a/tests/btrfs/192 b/tests/btrfs/192
> new file mode 100755
> index 00000000..db9bc40e
> --- /dev/null
> +++ b/tests/btrfs/192
> @@ -0,0 +1,156 @@
> +#! /bin/bash
> +# SPDX-License-Identifier: GPL-2.0
> +# Copyright (C) 2019 SUSE Linux Products GmbH. All Rights Reserved.
> +#
> +# FS QA Test 192
> +#
> +# Test btrfs consistency after each FUA for a workload with snapshot creation
> +# and removal
> +#
> +seq=`basename $0`
> +seqres=$RESULT_DIR/$seq
> +echo "QA output created by $seq"
> +
> +here=`pwd`
> +tmp=/tmp/$$
> +status=1 # failure is the default!
> +trap "_cleanup; exit \$status" 0 1 2 3 15
> +
> +_cleanup()
> +{
> + cd /
> + kill -q $pid1 &> /dev/null
> + kill -q $pid2 &> /dev/null
> + "$KILLALL_PROG" -q $FSSTRESS_PROG &> /dev/null
> + wait
> + _log_writes_cleanup &> /dev/null
> + rm -f $tmp.*
> +}
> +
> +# get standard environment, filters and checks
> +. ./common/rc
> +. ./common/filter
> +. ./common/attr
> +. ./common/dmlogwrites
> +
> +# remove previous $seqres.full before test
> +rm -f $seqres.full
> +
> +# real QA test starts here
> +
> +# Modify as appropriate.
> +_supported_fs btrfs
> +_supported_os Linux
> +
> +_require_command "$KILLALL_PROG" killall
> +_require_command "$BLKDISCARD_PROG" blkdiscard
> +_require_btrfs_fs_feature "no_holes"
> +_require_btrfs_mkfs_feature "no-holes"
> +_require_fsck_not_report_dirty_logs_as_error
> +_require_log_writes
> +_require_scratch
> +_require_attrs
> +
> +# To generate 3 level fs tree for 64K nodesize, we need 32768 xattr items.
> +# That will cause too many transactions, bumping replay check time
> +# from ~60s to ~300s. (VM alreayd using unsafe cache for the test devices)
> +# So here we skip non-4K page size system, in favor of a shorter default
> +# test time
> +if [ $(get_page_size) -ne 4096 ]; then
> + _notrun "This test doesn't support non-4K page size yet"
> +fi
> +
> +runtime=30
> +nr_cpus=$("$here/src/feature" -o)
> +# cap nr_cpus to 8 to avoid spending too much time on hosts with many cpus
> +if [ $nr_cpus -gt 8 ]; then
> + nr_cpus=8
> +fi
> +fsstress_args=$(_scale_fsstress_args -w -d $SCRATCH_MNT -n 99999 -p $nr_cpus \
> + $FSSTRESS_AVOID)
> +_log_writes_init $SCRATCH_DEV
> +
> +# Discard the whole devices so when some tree pointer is wrong, it won't point
> +# to some older valid tree blocks, so we can detect it.
> +$BLKDISCARD_PROG $LOGWRITES_DMDEV > /dev/null 2>&1
> +
> +# Workaround minor file extent discountinous.
> +# And use 4K nodesize to bump tree height.
> +_log_writes_mkfs -O no-holes -n 4k >> $seqres.full
> +_log_writes_mount
> +
> +$BTRFS_UTIL_PROG subvolume create $SCRATCH_MNT/src > /dev/null
> +mkdir -p $SCRATCH_MNT/snapshots
> +mkdir -p $SCRATCH_MNT/src/padding
> +
> +random_file()
> +{
> + local basedir=$1
> + echo "$basedir/$(ls $basedir | sort -R | tail -1)"
> +}
> +
> +snapshot_workload()
> +{
> + trap "wait; exit" SIGTERM
> +
> + local i=0
> + while true; do
> + $BTRFS_UTIL_PROG subvolume snapshot \
> + $SCRATCH_MNT/src $SCRATCH_MNT/snapshots/$i \
> + > /dev/null
> + # Do something small to make snapshots different
> + rm -f "$(random_file $SCRATCH_MNT/src/padding)"
> + rm -f "$(random_file $SCRATCH_MNT/src/padding)"
> + touch "$(random_file $SCRATCH_MNT/src/padding)"
> + touch "$SCRATCH_MNT/src/padding/random_$RANDOM"
> +
> + i=$(($i + 1))
> + sleep 1
> + done
> +}
> +
> +delete_workload()
> +{
> + trap "wait; exit" SIGTERM
> +
> + while true; do
> + sleep 2
> + $BTRFS_UTIL_PROG subvolume delete \
> + "$(random_file $SCRATCH_MNT/snapshots)" \
> + > /dev/null 2>&1
> + done
> +}
> +
> +xattr_value=$(printf '%0.sX' $(seq 1 3800))
> +
> +# Bumping tree height to level 2.
> +for ((i = 0; i < 64; i++)); do
> + touch "$SCRATCH_MNT/src/padding/$i"
> + $SETFATTR_PROG -n 'user.x1' -v $xattr_value \
> + "$SCRATCH_MNT/src/padding/$i"
> +done
> +
> +_log_writes_mark prepare
> +
> +snapshot_workload &
> +pid1=$!
> +delete_workload &
> +pid2=$!
> +
> +"$FSSTRESS_PROG" $fsstress_args > /dev/null &
> +sleep $runtime
> +
> +"$KILLALL_PROG" -q "$FSSTRESS_PROG" &> /dev/null
> +kill $pid1 &> /dev/null
> +kill $pid2 &> /dev/null
> +wait
> +_log_writes_unmount
> +_log_writes_remove
> +
> +_log_writes_fast_replay_check fua "$SCRATCH_DEV"
> +
> +echo "Silence is golden"
> +
> +# success, all done
> +status=0
> +exit
> diff --git a/tests/btrfs/192.out b/tests/btrfs/192.out
> new file mode 100644
> index 00000000..6779aa77
> --- /dev/null
> +++ b/tests/btrfs/192.out
> @@ -0,0 +1,2 @@
> +QA output created by 192
> +Silence is golden
> diff --git a/tests/btrfs/group b/tests/btrfs/group
> index 2474d43e..cab10d19 100644
> --- a/tests/btrfs/group
> +++ b/tests/btrfs/group
> @@ -194,3 +194,4 @@
> 189 auto quick send clone
> 190 auto quick replay balance qgroup
> 191 auto quick send dedupe
> +192 auto replay snapshot stress
> --
> 2.22.0
>
