Re: [PATCH] xfs/842: verify CoW after exchangerange with FILE1_WRITTEN on shared extents

On 2026/7/28 22:59, Darrick J. Wong Wrote:
> On Tue, Jul 28, 2026 at 03:16:20PM +0800, Lin Jiapeng wrote:
>> When a full-file exchangerange is requested with FILE1_WRITTEN for a
>> fully sparse donor file, every mapping pair is skipped, so no extents
>> actually move.  However, the kernel decides to swap the reflink inode
>> flag based on the request geometry alone, so the target file used to
>> lose its reflink flag while still owning shared extents.  A subsequent
>> write then took the non-reflink write path and modified the shared
>> physical blocks in place, silently corrupting the other file sharing
>> them:
>>
>>   --- tests/xfs/842.out
>>   +++ results/xfs/842.out.bad
>>   @@ -1,2 +1,3 @@
>>    QA output created by 842
>>   +orig changed: md5 e6065c4aa2ab1603008fc18410f579d4 -> 32c5189478e6bafa1cc76423f06c88f2 (write hit shared blocks in place)
>>    Silence is golden
>>
>> This test clones a file so that both share extents, swaps the clone
>> against a sparse donor with FILE1_WRITTEN, writes to the clone, and
>> verifies after a mount cycle that the original file's data is intact,
>> i.e. the write was redirected through CoW.
>>
>> The kernel fix "xfs: fix exchange-range reflink flag clearing issue
>> with INO1_WRITTEN" refuses the reflink flag exchange whenever
>> FILE1_WRITTEN is requested, which makes this test pass.
>>
>> Reported-by: Lin Jiapeng(TencentOS Red Team) <[email protected]>
>> Link: https://lore.kernel.org/r/[email protected]
>> Suggested-by: Christoph Hellwig <[email protected]>
>> Signed-off-by: Lin Jiapeng <[email protected]>
>> ---
>>   tests/xfs/842     | 55 +++++++++++++++++++++++++++++++++++++++++++++++
>>   tests/xfs/842.out |  2 ++
>>   2 files changed, 57 insertions(+)
>>   create mode 100755 tests/xfs/842
>>   create mode 100644 tests/xfs/842.out
>>
>> diff --git a/tests/xfs/842 b/tests/xfs/842
>> new file mode 100755
>> index 0000000..6d5bcab
>> --- /dev/null
>> +++ b/tests/xfs/842
>> @@ -0,0 +1,55 @@
>> +#! /bin/bash
>> +# SPDX-License-Identifier: GPL-2.0-or-later
>> +# Copyright (c) 2026 Tencent.  All Rights Reserved.
>> +#
>> +# FS QA Test No. 842
>> +#
>> +# Make sure that a full-file exchangerange under FILE1_WRITTEN does not strip
>> +# the reflink flag from a file that still owns shared extents.  When the
>> +# donor file is fully sparse, every mapping pair is skipped, so no extents
>> +# actually move; the target file must keep its reflink flag, and a later
>> +# write must go through CoW instead of modifying the shared blocks in place.
>> +
>> +. ./common/preamble
>> +_begin_fstest auto quick fiexchange
>> +
>> +# Import common functions.
>> +. ./common/filter
>> +. ./common/reflink
>> +
>> +_require_xfs_io_command exchangerange
>> +_require_scratch_reflink
>> +_require_scratch
>> +
>> +_scratch_mkfs >> $seqres.full
>> +_scratch_mount
>> +
>> +# Create the original file with a known pattern and clone it, so that both
>> +# files share the same extents.
>> +_pwrite_byte 0x41 0 1m $SCRATCH_MNT/orig >> $seqres.full
>> +_reflink $SCRATCH_MNT/orig $SCRATCH_MNT/clone >> $seqres.full
>> +
>> +# Create a fully sparse donor file of the same size.
>> +$XFS_IO_PROG -f -c 'truncate 1m' $SCRATCH_MNT/donor
> 
> Hehe, this is exactly the same sequence as the testcase I wrote
> yesterday afternoon, so the review is easy;
> 
> Reviewed-by: "Darrick J. Wong" <[email protected]>
> 
> (For extra clarity I also had done:
> 
> # Check reflink flags
> _scratch_unmount
> _scratch_xfs_db -c "path /orig" -c 'print' -c "path /donor" -c 'print' | grep reflink
> _scratch_mount
> 
> for direct confirmation of the incorrect inode flag state, but the
> md5sum failure suffices to prove the defect.)
> 
> --D
> 
>> +
>> +md5_before=$(md5sum $SCRATCH_MNT/orig | awk '{print $1}')
>> +
>> +# Swap the clone against the sparse donor, claiming the donor is fully
>> +# written (-w).  Every donor mapping is a hole, so all pairs are skipped
>> +# and the clone keeps its shared extents in place.
>> +$XFS_IO_PROG -c "exchangerange -f -w $SCRATCH_MNT/donor" $SCRATCH_MNT/clone \
>> +	>> $seqres.full
>> +
>> +# Overwrite part of the clone.  With the reflink flag correctly retained,
>> +# this must go through CoW and leave the shared blocks of orig untouched.
>> +_pwrite_byte 0x42 0 64k $SCRATCH_MNT/clone >> $seqres.full
>> +_scratch_cycle_mount
>> +
>> +md5_after=$(md5sum $SCRATCH_MNT/orig | awk '{print $1}')
>> +
>> +test "$md5_before" != "$md5_after" && \
>> +	echo "orig changed: md5 $md5_before -> $md5_after (write hit shared blocks in place)"
>> +
>> +echo Silence is golden
>> +status=0
>> +exit
>> diff --git a/tests/xfs/842.out b/tests/xfs/842.out
>> new file mode 100644
>> index 0000000..643d052
>> --- /dev/null
>> +++ b/tests/xfs/842.out
>> @@ -0,0 +1,2 @@
>> +QA output created by 842
>> +Silence is golden
>> -- 
>> 2.50.1 (Apple Git-155)
>>
>>
Good idea! I will add direct confirmation of the incorrect inode flag 
state in Patch v2.