apfstests/tests/xfs/444

#! /bin/bash
# SPDX-License-Identifier: GPL-2.0
# Copyright (c) 2018 Oracle, Inc.
#
# FS QA Test No. 444
#
# Make sure XFS can fix a v5 AGFL that wraps over the last block.
# Refer to commit 96f859d52bcb ("libxfs: pack the agfl header structure so
# XFS_AGFL_SIZE is correct") for details on the original on-disk format error
# and the patch "xfs: detect agfl count corruption and reset agfl") for details
# about the fix.
#
seq=`basename $0`
seqres=$RESULT_DIR/$seq
echo "QA output created by $seq"

here=`pwd`
tmp=/tmp/$$
status=1
trap "_cleanup; rm -f $tmp.*; exit \$status" 0 1 2 3 15

_cleanup()
{
	cd /
	rm -f $tmp.*
}

rm -f $seqres.full

# get standard environment, filters and checks
. ./common/rc
. ./common/filter

# real QA test starts here
_supported_fs xfs

_require_check_dmesg
_require_scratch
_require_test_program "punch-alternating"
_require_xfs_io_command "falloc"
_require_xfs_db_write_array

# This is only a v5 filesystem problem
_require_scratch_xfs_crc

# Disable the scratch rt device to avoid test failures relating to the rt
# bitmap consuming free space in our small data device and throwing off the
# filestreams allocator.
unset SCRATCH_RTDEV

mount_loop() {
	if ! _try_scratch_mount >> $seqres.full 2>&1; then
		echo "scratch mount failed" >> $seqres.full
		return
	fi

	# Trigger agfl fixing by fragmenting free space enough to cause
	# a bnobt split
	blksz=$(_get_file_block_size ${SCRATCH_MNT})
	bno_maxrecs=$(( blksz / 8 ))
	filesz=$((bno_maxrecs * 3 * blksz))
	rm -rf $SCRATCH_MNT/a
	$XFS_IO_PROG -f -c "falloc 0 $filesz" $SCRATCH_MNT/a >> $seqres.full 2>&1
	test -e $SCRATCH_MNT/a && $here/src/punch-alternating $SCRATCH_MNT/a
	rm -rf $SCRATCH_MNT/a

	_scratch_unmount 2>&1 | _filter_scratch
}

dump_ag0() {
	_scratch_xfs_db -c 'sb 0' -c 'p' -c 'agf 0' -c 'p' -c 'agfl 0' -c 'p'
}

runtest() {
	cmd="$1"

	# Format filesystem
	echo "TEST $cmd" | tee /dev/ttyprintk
	echo "TEST $cmd" >> $seqres.full
	_scratch_mkfs >> $seqres.full

	# Record what was here before
	echo "FS BEFORE" >> $seqres.full
	dump_ag0 > $tmp.before
	cat $tmp.before >> $seqres.full

	sectsize=$(_scratch_xfs_get_metadata_field "sectsize" "sb 0")
	flfirst=$(_scratch_xfs_get_metadata_field "flfirst" "agf 0")
	fllast=$(_scratch_xfs_get_metadata_field "fllast" "agf 0")
	flcount=$(_scratch_xfs_get_metadata_field "flcount" "agf 0")

	# Due to a padding bug in the original v5 struct xfs_agfl,
	# XFS_AGFL_SIZE could be 36 on 32-bit or 40 on 64-bit.  On a system
	# with 512b sectors, this means that the AGFL length could be
	# ((512 - 36) / 4) = 119 entries on 32-bit or ((512 - 40) / 4) = 118
	# entries on 64-bit.
	#
	# We now have code to figure out if the AGFL list wraps incorrectly
	# according to the kernel's agfl size and fix it by resetting the agfl
	# to zero length.  Mutate ag 0's agfl to be in various configurations
	# and see if we can trigger the reset.
	#
	# Don't hardcode the numbers, calculate them.

	# Have to have at least three agfl items to test full wrap
	test "$flcount" -ge 3 || _notrun "insufficient agfl flcount"

	# mkfs should be able to make us a nice neat flfirst < fllast setup
	test "$flfirst" -lt "$fllast" || _notrun "fresh agfl already wrapped?"

	bad_agfl_size=$(( (sectsize - 40) / 4 ))
	good_agfl_size=$(( (sectsize - 36) / 4 ))
	agfl_size=
	case "$1" in
	"fix_end")	# fllast points to the end w/ 40-byte padding
		new_flfirst=$(( bad_agfl_size - flcount ))
		agfl_size=$bad_agfl_size;;
	"fix_start")	# flfirst points to the end w/ 40-byte padding
		new_flfirst=$(( bad_agfl_size - 1))
		agfl_size=$bad_agfl_size;;
	"fix_wrap")	# list wraps around end w/ 40-byte padding
		new_flfirst=$(( bad_agfl_size - (flcount / 2) ))
		agfl_size=$bad_agfl_size;;
	"start_zero")	# flfirst points to the start
		new_flfirst=0
		agfl_size=$good_agfl_size;;
	"good_end")	# fllast points to the end w/ 36-byte padding
		new_flfirst=$(( good_agfl_size - flcount ))
		agfl_size=$good_agfl_size;;
	"good_start")	# flfirst points to the end w/ 36-byte padding
		new_flfirst=$(( good_agfl_size - 1 ))
		agfl_size=$good_agfl_size;;
	"good_wrap")	# list wraps around end w/ 36-byte padding
		new_flfirst=$(( good_agfl_size - (flcount / 2) ))
		agfl_size=$good_agfl_size;;
	"bad_start")	# flfirst points off the end
		new_flfirst=$good_agfl_size
		agfl_size=$good_agfl_size;;
	"no_move")	# whatever mkfs formats (flfirst points to start)
		new_flfirst=$flfirst
		agfl_size=$good_agfl_size;;
	"simple_move")	# move list arbitrarily
		new_flfirst=$((fllast + 1))
		agfl_size=$good_agfl_size;;
	*)
		_fail "Internal test error";;
	esac
	new_fllast=$(( (new_flfirst + flcount - 1) % agfl_size ))

	# Log what we're doing...
	cat >> $seqres.full << ENDL
sector size: $sectsize
bad_agfl_size: $bad_agfl_size [0 - $((bad_agfl_size - 1))]
good_agfl_size: $good_agfl_size [0 - $((good_agfl_size - 1))]
agfl_size: $agfl_size
flfirst: $flfirst
fllast: $fllast
flcount: $flcount
new_flfirst: $new_flfirst
new_fllast: $new_fllast
ENDL

	# Remap the agfl blocks
	echo "$((good_agfl_size - 1)) 0xffffffff" > $tmp.remap
	seq "$flfirst" "$fllast" | while read f; do
		list_pos=$((f - flfirst))
		dest_pos=$(( (new_flfirst + list_pos) % agfl_size ))
		bno=$(_scratch_xfs_get_metadata_field "bno[$f]" "agfl 0")
		echo "$dest_pos $bno" >> $tmp.remap
	done

	cat $tmp.remap | while read dest_pos bno junk; do
		_scratch_xfs_set_metadata_field "bno[$dest_pos]" "$bno" \
				"agfl 0" >> $seqres.full
	done

	# Set new flfirst/fllast
	_scratch_xfs_set_metadata_field "fllast" "$new_fllast" \
			"agf 0" >> $seqres.full
	_scratch_xfs_set_metadata_field "flfirst" "$new_flfirst" \
			"agf 0" >> $seqres.full

	echo "FS AFTER" >> $seqres.full
	dump_ag0 > $tmp.corrupt 2> /dev/null
	diff -u $tmp.before $tmp.corrupt >> $seqres.full

	# Mount and see what happens
	mount_loop

	# Did we end up with a non-wrapped list?
	flfirst=$(_scratch_xfs_get_metadata_field "flfirst" "agf 0" 2>/dev/null)
	fllast=$(_scratch_xfs_get_metadata_field "fllast" "agf 0" 2>/dev/null)
	echo "flfirst=${flfirst} fllast=${fllast}" >> $seqres.full
	if [ "${flfirst}" -ge "$((good_agfl_size - 1))" ]; then
		echo "ASSERT flfirst < good_agfl_size - 1" | tee -a $seqres.full
	fi
	if [ "${fllast}" -ge "$((good_agfl_size - 1))" ]; then
		echo "ASSERT fllast < good_agfl_size - 1" | tee -a $seqres.full
	fi
	if [ "${flfirst}" -ge "${fllast}" ]; then
		echo "ASSERT flfirst < fllast" | tee -a $seqres.full
	fi

	echo "FS MOUNTLOOP" >> $seqres.full
	dump_ag0 > $tmp.mountloop 2> /dev/null
	diff -u $tmp.corrupt $tmp.mountloop >> $seqres.full

	# Let's see what repair thinks
	echo "REPAIR" >> $seqres.full
	_scratch_xfs_repair >> $seqres.full 2>&1

	echo "FS REPAIR" >> $seqres.full
	dump_ag0 > $tmp.repair 2> /dev/null
	diff -u $tmp.mountloop $tmp.repair >> $seqres.full

	# Exercise the filesystem again to make sure there aren't any lasting
	# ill effects from either the agfl reset or the recommended subsequent
	# repair run.
	mount_loop

	echo "FS REMOUNT" >> $seqres.full
	dump_ag0 > $tmp.remount 2> /dev/null
	diff -u $tmp.repair $tmp.remount >> $seqres.full
}

runtest fix_end
runtest fix_start
runtest fix_wrap
runtest start_zero
runtest good_end
runtest good_start
runtest good_wrap
runtest bad_start
runtest no_move
runtest simple_move

# Did we get the kernel warning too?
warn_str='WARNING: Reset corrupted AGFL'
_check_dmesg_for "${warn_str}" || echo "Missing dmesg string \"${warn_str}\"."

# Now run the regular dmesg check, filtering out the agfl warning
filter_agfl_reset_printk() {
	grep -v "${warn_str}"
}
_check_dmesg filter_agfl_reset_printk

status=0
exit