Re: [PATCH] xfs: test the deadlock between the AGI and AGF with RENAME_WHITEOUT

[kaixuxia <xiakaixu1987@gmail.com>]

On 2019/9/8 20:39, Eryu Guan wrote:
> On Wed, Sep 04, 2019 at 06:01:27PM +0800, kaixuxia wrote:
>> There is ABBA deadlock bug between the AGI and AGF when performing
>> rename() with RENAME_WHITEOUT flag, and add this testcase to make
>> sure the rename() call works well.
>>
>> Signed-off-by: kaixuxia <kaixuxia@tencent.com>
>> ---
>>  tests/xfs/512     | 100 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
>>  tests/xfs/512.out |   2 ++
>>  tests/xfs/group   |   1 +
>>  3 files changed, 103 insertions(+)
>>  create mode 100755 tests/xfs/512
>>  create mode 100644 tests/xfs/512.out
>>
>> diff --git a/tests/xfs/512 b/tests/xfs/512
>> new file mode 100755
>> index 0000000..0e95fb7
>> --- /dev/null
>> +++ b/tests/xfs/512
>> @@ -0,0 +1,100 @@
>> +#! /bin/bash
>> +# SPDX-License-Identifier: GPL-2.0
>> +# Copyright (c) 2019 Tencent.  All Rights Reserved.
>> +#
>> +# FS QA Test 512
>> +#
>> +# Test the ABBA deadlock case between the AGI and AGF When performing
>> +# rename operation with RENAME_WHITEOUT flag.
>> +#
>> +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
>> +
>> +rm -f $seqres.full
>> +
>> +# real QA test starts here
>> +_supported_fs xfs
>> +_supported_os Linux
>> +_require_scratch
> 
> Only _require_scratch_nocheck is suffiecient.

Yeah, will fix it.
> 
>> +
>> +# Single AG will cause default xfs_repair to fail. This test need a
>> +# single AG fs, so ignore the check.
>> +_require_scratch_nocheck
> 
> Also need to 
> 
> . ./common/rename
> ...
> 
> _requires_renameat2
> 
> Also, this test requires RENAME_WHITEOUT, I'd suggest enhance
> src/renameat2.c to support check for if a given rename flag is supported
> by kernel, and refactor the checks in generic/02[45] and generic/078
> into _requires_renameat2 to use the new functionality. e.g.
> 
> # without option, behavior stays unchanged, check for renameat2 syscall
> # support
> _requires_renameat2
> 
> # check if renameat2 upports RENAME_WHITEOUT flag
> _requires_renameat2 whiteout
> 
> # check if renameat2 upports RENAME_EXCHANGE flag
> _requires_renameat2 exchange
> 

Right, this testcase requires RENAME_WHITEOUT, it is necessary that check
whether the given rename flag is supported by kernel. I will try to refactor
the corresponding tests and functions.
>> +
>> +prepare_file()
>> +{
>> +	# create many small files for the rename with RENAME_WHITEOUT
>> +	i=0
>> +	while [ $i -le $files ]; do
>> +		file=$SCRATCH_MNT/f$i
>> +		$XFS_IO_PROG -f -d -c 'pwrite -b 4k 0 4k' $file >/dev/null 2>&1
>> +		let i=$i+1
> 
> Creating 250000 4k files take a long time. Does file content really
> matters? I guess racing RENAME_WHITEOUT with file creation is enough, is
> it possible to just create many empty files? e.g.
> 
> 		echo > $file
> 
> this saves a lot time.

Yeah, create empty files is better.
> 
>> +	done
>> +}
>> +
>> +rename_whiteout()
>> +{
>> +	# create the rename targetdir
>> +	renamedir=$SCRATCH_MNT/renamedir
>> +	mkdir $renamedir
>> +
>> +	# just get a random long name...
>> +	longnamepre=FFFsafdsagafsadfagasdjfalskdgakdlsglkasdg
> 
> Better to explain why long file name is required.
> 
Will add the explanation. The long file name will increase the possibility
that target_dp allocate new blocks(acquire the AGF lock) to store the
filename. 
>> +
>> +	# now try to do rename with RENAME_WHITEOUT flag
>> +	i=0
>> +	while [ $i -le $files ]; do
>> +		src/renameat2 -w $SCRATCH_MNT/f$i $renamedir/$longnamepre$i >/dev/null 2>&1
>> +		let i=$i+1
>> +	done
>> +}
>> +
>> +create_file()
>> +{
>> +	# create the targetdir
>> +	createdir=$SCRATCH_MNT/createdir
>> +	mkdir $createdir
>> +
>> +	# try to create file at the same time to hit the deadlock
>> +	i=0
>> +	while [ $i -le $files ]; do
>> +		file=$createdir/f$i
>> +		$XFS_IO_PROG -f -d -c 'pwrite -b 4k 0 4k' $file >/dev/null 2>&1
>> +		let i=$i+1
>> +	done
> 
> Same here, does creating empty files work?

Yeah.
> 
>> +}
>> +
>> +_scratch_mkfs_xfs -bsize=512 -dagcount=1 >> $seqres.full 2>&1 ||
> 
> This doesn't work because crc is on by default, as crc requires minimum
> 1k block size. Is 512 block size really needed?
> 
>> +	_fail "mkfs failed"
>> +_scratch_mount
>> +
>> +files=250000
> 
> If we could reduce file number to create, test could run faster as well.
> Running test with less than 250000 files couldn't reproduce the
> deadlock?

I used this magic number in my local test environment and reproduced
the deadlock every time. Creating empty files would reduce the 
run time...

Thanks.
> 
> Thanks,
> Eryu
> 
>> +
>> +prepare_file
>> +rename_whiteout &
>> +create_file &
>> +
>> +wait
>> +echo Silence is golden
>> +
>> +# Failure comes in the form of a deadlock.
>> +
>> +# success, all done
>> +status=0
>> +exit
>> diff --git a/tests/xfs/512.out b/tests/xfs/512.out
>> new file mode 100644
>> index 0000000..0aabdef
>> --- /dev/null
>> +++ b/tests/xfs/512.out
>> @@ -0,0 +1,2 @@
>> +QA output created by 512
>> +Silence is golden
>> diff --git a/tests/xfs/group b/tests/xfs/group
>> index a7ad300..ed250d6 100644
>> --- a/tests/xfs/group
>> +++ b/tests/xfs/group
>> @@ -509,3 +509,4 @@
>>  509 auto ioctl
>>  510 auto ioctl quick
>>  511 auto quick quota
>> +512 auto rename
>> -- 
>> 1.8.3.1
>>
>> -- 
>> kaixuxia

-- 
kaixuxia