Re: [PATCH V4 07/11] xfs: Check for extent overflow when writing to unwritten extent

[Chandan Babu R <chandanrlinux@gmail.com>]

On 03 Mar 2021 at 23:35, Darrick J. Wong wrote:
> On Mon, Jan 18, 2021 at 11:50:18AM +0530, Chandan Babu R wrote:
>> This test verifies that XFS does not cause inode fork's extent count to
>> overflow when writing to an unwritten extent.
>>
>> Signed-off-by: Chandan Babu R <chandanrlinux@gmail.com>
>> ---
>>  tests/xfs/527     | 89 +++++++++++++++++++++++++++++++++++++++++++++++
>>  tests/xfs/527.out | 11 ++++++
>>  tests/xfs/group   |  1 +
>>  3 files changed, 101 insertions(+)
>>  create mode 100755 tests/xfs/527
>>  create mode 100644 tests/xfs/527.out
>>
>> diff --git a/tests/xfs/527 b/tests/xfs/527
>> new file mode 100755
>> index 00000000..cd67bce4
>> --- /dev/null
>> +++ b/tests/xfs/527
>> @@ -0,0 +1,89 @@
>> +#! /bin/bash
>> +# SPDX-License-Identifier: GPL-2.0
>> +# Copyright (c) 2020 Chandan Babu R.  All Rights Reserved.
>> +#
>> +# FS QA Test 527
>> +#
>> +# Verify that XFS does not cause inode fork's extent count to overflow when
>> +# writing to an unwritten extent.
>> +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 /
>> +	rm -f $tmp.*
>> +}
>> +
>> +# get standard environment, filters and checks
>> +. ./common/rc
>> +. ./common/filter
>> +. ./common/inject
>> +
>> +# remove previous $seqres.full before test
>> +rm -f $seqres.full
>> +
>> +# real QA test starts here
>> +
>> +_supported_fs xfs
>> +_require_scratch
>> +_require_xfs_debug
>> +_require_xfs_io_command "falloc"
>> +_require_xfs_io_error_injection "reduce_max_iextents"
>> +
>> +echo "Format and mount fs"
>> +_scratch_mkfs_sized $((1024 * 1024 * 1024)) >> $seqres.full
>> +_scratch_mount >> $seqres.full
>> +
>> +bsize=$(_get_file_block_size $SCRATCH_MNT)
>> +
>> +testfile=${SCRATCH_MNT}/testfile
>> +
>> +echo "Inject reduce_max_iextents error tag"
>> +_scratch_inject_error reduce_max_iextents 1
>> +
>> +nr_blks=15
>> +
>> +for io in Buffered Direct; do
>
> # First test buffered writes, then direct writes.
> for $xfs_io_flag in "" "-d"; do ?

The keywords "Buffered" and "Direct" are also being used in the log
messages printed ...

>
>> +	echo "* $io write to unwritten extent"

... For example, In the above line, $io would be replaced by one of "Buffered"
and "Direct" ...

>> +
>> +	echo "Fallocate $nr_blks blocks"
>> +	$XFS_IO_PROG -f -c "falloc 0 $((nr_blks * bsize))" $testfile >> $seqres.full
>> +
>> +	if [[ $io == "Buffered" ]]; then
>> +		xfs_io_flag=""
>> +	else
>> +		xfs_io_flag="-d"
>> +	fi
>
> ...because then you can skip this part.
>
>> +
>> +	echo "$io write to every other block of fallocated space"

... Same logic is used above.

>> +	for i in $(seq 1 2 $((nr_blks - 1))); do
>> +		$XFS_IO_PROG -f -s $xfs_io_flag -c "pwrite $((i * bsize)) $bsize" \
>> +		       $testfile >> $seqres.full 2>&1
>> +		[[ $? != 0 ]] && break
>> +	done
>> +
>> +	echo "Verify \$testfile's extent count"
>> +	nextents=$($XFS_IO_PROG -c 'stat' $testfile | grep nextents)
>> +	nextents=${nextents##fsxattr.nextents = }
>> +	if (( $nextents > 10 )); then
>> +		echo "Extent count overflow check failed: nextents = $nextents"
>> +		exit 1
>> +	fi
>> +
>> +	rm $testfile
>> +done
>> +
>> +# super_block->s_wb_err will have a newer seq value when compared to "/"'s
>> +# file->f_sb_err. Consume it here so that xfs_scrub can does not error out.
>> +$XFS_IO_PROG -c syncfs $SCRATCH_MNT >> $seqres.full 2>&1
>
> I wonder, should _check_xfs_filesystem should syncfs to clear old EIOs
> before running xfs_scrub?  I occasionally see this pop up on
> generic/204.

I agree with the idea of including a call to syncfs in
_check_xfs_filesystem. An error message can be logged if syncfs returns a
failure status. This will also prevent xfs_scrub from aborting and help in
providing information about any corrupt on-disk data structures.

I will include this in the next version of the patchset.

--
chandan