On Mon, Aug 16, 2021 at 09:16:16PM -0700, Darrick J. Wong wrote:

On Tue, Aug 17, 2021 at 11:16:49AM +0800, Eryu Guan wrote:

quoted

On Mon, Aug 16, 2021 at 09:35:24AM -0700, Darrick J. Wong wrote:

quoted

On Mon, Aug 16, 2021 at 12:28:20AM +0800, Eryu Guan wrote:

quoted

On Tue, Jul 27, 2021 at 05:10:30PM -0700, Darrick J. Wong wrote:

quoted

From: Darrick J. Wong <djwong@kernel.org>

generic/475, but we're running fsstress on a disk image inside the
scratch filesystem

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
---
 tests/generic/725     |  136 +++++++++++++++++++++++++++++++++++++++++++++++++
 tests/generic/725.out |    2 +
 2 files changed, 138 insertions(+)
 create mode 100755 tests/generic/725
 create mode 100644 tests/generic/725.out

diff --git a/tests/generic/725 b/tests/generic/725
new file mode 100755
index 00000000..f43bcb37
--- /dev/null
+++ b/tests/generic/725

@@ -0,0 +1,136 @@
+#! /bin/bash
+# SPDX-License-Identifier: GPL-2.0
+# Copyright (c) 2021 Oracle, Inc.  All Rights Reserved.
+#
+# FS QA Test No. 725
+#
+# Test nested log recovery with repeated (simulated) disk failures.  We kick
+# off fsstress on a loopback filesystem mounted on the scratch fs, then switch
+# out the underlying scratch device with dm-error to see what happens when the
+# disk goes down.  Having taken down both fses in this manner, remount them and
+# repeat.  This test simulates VM hosts crashing to try to shake out CoW bugs
+# in writeback on the host that cause VM guests to fail to recover.

It currently fails for me on btrfs, the loop mount failed in 2nd
iteration, seems like a bug in btrfs.

Yep.  Until recently (aka the Big Xfs Log Recovery Bughunt of 2021) it
wouldn't pass xfs either. :/

quoted

quoted

+#
+. ./common/preamble
+_begin_fstest shutdown auto log metadata eio
+
+_cleanup()
+{
+	cd /
+	$KILLALL_PROG -9 fsstress > /dev/null 2>&1
+	wait
+	if [ -n "$loopmnt" ]; then
+		umount $loopmnt 2>/dev/null

$UMOUNT_PROG

quoted

+		rm -r -f $loopmnt
+	fi
+	rm -f $tmp.*
+	_dmerror_unmount
+	_dmerror_cleanup
+}
+
+# Import common functions.
+. ./common/dmerror
+. ./common/reflink
+
+# Modify as appropriate.
+_supported_fs generic
+
+_require_scratch_reflink
+_require_cp_reflink
+_require_dm_target error
+_require_command "$KILLALL_PROG" "killall"
+
+echo "Silence is golden."
+
+_scratch_mkfs >> $seqres.full 2>&1
+_require_metadata_journaling $SCRATCH_DEV
+_dmerror_init
+_dmerror_mount
+
+# Create a fs image consuming 1/3 of the scratch fs
+scratch_freesp_bytes=$(stat -f -c '%a * %S' $SCRATCH_MNT | bc)

_get_available_space $SCRATCH_MNT ?

quoted

+loopimg_bytes=$((scratch_freesp_bytes / 3))
+
+loopimg=$SCRATCH_MNT/testfs
+truncate -s $loopimg_bytes $loopimg
+_mkfs_dev $loopimg
+
+loopmnt=$tmp.mount
+mkdir -p $loopmnt
+
+scratch_aliveflag=$tmp.runsnap
+snap_aliveflag=$tmp.snapping
+
+snap_loop_fs() {
+	touch "$snap_aliveflag"
+	while [ -e "$scratch_aliveflag" ]; do
+		rm -f $loopimg.a
+		_cp_reflink $loopimg $loopimg.a
+		sleep 1
+	done
+	rm -f "$snap_aliveflag"
+}
+
+fsstress=($FSSTRESS_PROG $FSSTRESS_AVOID -d "$loopmnt" -n 999999 -p "$((LOAD_FACTOR * 4))")
+
+for i in $(seq 1 $((25 * TIME_FACTOR)) ); do
+	touch $scratch_aliveflag
+	snap_loop_fs >> $seqres.full 2>&1 &
+
+	if ! _mount $loopimg $loopmnt -o loop; then
+		rm -f $scratch_aliveflag
+		_fail "loop mount failed"

I found it a bit easier to debug if print $i here.

Ok, I'll change it to "loop $i mount failed".

quoted

quoted

+		break
+	fi
+
+	("${fsstress[@]}" >> $seqres.full &) > /dev/null 2>&1
+
+	# purposely include 0 second sleeps to test shutdown immediately after
+	# recovery
+	sleep $((RANDOM % (3 * TIME_FACTOR) ))
+	rm -f $scratch_aliveflag
+
+	# This test aims to simulate sudden disk failure, which means that we
+	# do not want to quiesce the filesystem or otherwise give it a chance
+	# to flush its logs.  Therefore we want to call dmsetup with the
+	# --nolockfs parameter; to make this happen we must call the load
+	# error table helper *without* 'lockfs'.
+	_dmerror_load_error_table
+
+	ps -e | grep fsstress > /dev/null 2>&1
+	while [ $? -eq 0 ]; do
+		$KILLALL_PROG -9 fsstress > /dev/null 2>&1
+		wait > /dev/null 2>&1
+		ps -e | grep fsstress > /dev/null 2>&1
+	done
+	for ((i = 0; i < 10; i++)); do
+		test -e "$snap_aliveflag" || break
+		sleep 1
+	done
+
+	# Mount again to replay log after loading working table, so we have a
+	# consistent XFS after test.

This is a generic test, fix the XFS specific comments?

Oops.  "...a consistent fs after test."

quoted

quoted

+	$UMOUNT_PROG $loopmnt
+	_dmerror_unmount || _fail "unmount failed"
+	_dmerror_load_working_table
+	if ! _dmerror_mount; then
+		dmsetup table | tee -a /dev/ttyprintk
+		lsblk | tee -a /dev/ttyprintk
+		$XFS_METADUMP_PROG -a -g -o $DMERROR_DEV $seqres.dmfail.md

Above logs all should go to $seqres.full ?

Oops, yeah.  I'll remove them since I was only using them to check the
system state.

quoted

And $XFS_METADUMP_PROG is not suitable for a generic test.

I'll create _metadump_dev so that this at least works for the two
filesystems for which we have dump creation helpers (ext* and xfs).

Sounds great!

quoted

quoted

quoted

+		_fail "mount failed"
+	fi
+done
+
+# Make sure the fs image file is ok
+if [ -f "$loopimg" ]; then
+	if _mount $loopimg $loopmnt -o loop; then
+		$UMOUNT_PROG $loopmnt &> /dev/null
+	else
+		echo "final loop mount failed"
+	fi
+	_check_xfs_filesystem $loopimg none none

Same here, use _check_scratch_fs?

$loopimg is a file within the scratch fs.

_check_scratch_fs can take dev as argument, and default to $SCRATCH_DEV,
I think that works in this case?

It could be made to work with a large enough crowbar, but that's
seriously overkill because $loopimg is a file *within* the scratch
filesystem.  The $loopimg fs gets formatted without the
SCRATCH_LOGDEV/SCRATCH_RTDEV options (because it is not itself the
scratch filesystem), which means that in order to (ab)use
_check_scratch_fs to do the same thing as _check_xfs_filesystem, you
have to exclude those options.  So yes, this:

	SCRATCH_RTDEV= SCRATCH_LOGDEV= _check_scratch_fs $loopimg

is the equivalent of this:

	_check_xfs_filesystem $loopimg none none

But the first is longer and pointless.

...and now that it's morning and I've had coffee again, I now understand
what you're actually asking, which is "Don't use _foo_xfs* functions in
a generic test!", not "rototill in this helper for stylistic reasons".

Judging from my immediate defensive reaction, I've clearly been worn
down by all the bikeshedding the past year.  Have they resumed flights
to Nobikeshed Island?

Anyway, I'll go fix that.  Thank you for catching the mistake.

--D

--D

quoted

Thanks,
Eryu