Re: [PATCH 4/6] xfs/128: try to force file allocation behavior

["Darrick J. Wong" <djwong@kernel.org>]

On Sat, Oct 08, 2022 at 07:11:02PM +0800, Zorro Lang wrote:
> On Wed, Oct 05, 2022 at 03:30:55PM -0700, Darrick J. Wong wrote:
> > From: Darrick J. Wong <djwong@kernel.org>
> > 
> > Over the years, I've noticed that this test occasionally fails when I've
> > programmed the allocator to hand out the minimum amount of space with
> > each allocation or if extent size hints are enabled:
> > 
> > --- /tmp/fstests/tests/xfs/128.out      2022-09-01 15:09:11.506679341 -0700
> > +++ /var/tmp/fstests/xfs/128.out.bad    2022-10-04 17:32:50.992000000 -0700
> > @@ -20,7 +21,9 @@
> >  56ed2f712c91e035adeeb26ed105a982  SCRATCH_MNT/test-128/file3
> >  b81534f439aac5c34ce3ed60a03eba70  SCRATCH_MNT/test-128/file4
> >  Check files
> >  free blocks after creating some reflink copies is in range
> >  free blocks after CoW some reflink copies is in range
> > -free blocks after defragging all reflink copies is in range
> > -free blocks after all tests is in range
> > +free blocks after defragging all reflink copies has value of 8620027
> > +free blocks after defragging all reflink copies is NOT in range 8651819 .. 8652139
> > +free blocks after all tests has value of 8620027
> > +free blocks after all tests is NOT in range 8651867 .. 8652187
> > 
> > It turns out that under the right circumstances, the _pwrite_byte at the
> > start of this test will end up allocating two extents to file1.  This
> > almost never happens when delalloc is enabled or when the extent size is
> > large, and is more prone to happening if the extent size is > 1FSB but
> > small, the allocator hands out small allocations, or if writeback shoots
> > down pages in random order.
> > 
> > When file1 gets more than 1 extent, problems start to happen.  The free
> > space accounting checks at the end of the test assume that file1 and
> > file4 still share the same space at the end of the test.  This
> > definitely happens if file1 gets one extent (since fsr ignores
> > single-extent files), but if there's more than 1, fsr will try to
> > defragment it.  If fsr succeeds in copying the file contents to a temp
> > file with fewer extents than the source file, it will switch the
> > contents, but unsharing the contents in the process.  This cause the
> > free space to be lower than expected, and the test fails.
> > 
> > Resolve this situation by preallocating space beforehand to try to set
> > up file1 with a single space extent.  If the test fails and we got more
> > than one extent, note that in the output.
> > 
> > Signed-off-by: Darrick J. Wong <djwong@kernel.org>
> > ---
> 
> Good to me,
> Reviewed-by: Zorro Lang <zlang@redhat.com>
> 
> >  tests/xfs/128 |   34 ++++++++++++++++++++++++++++++----
> >  1 file changed, 30 insertions(+), 4 deletions(-)
> > 
> > 
> > diff --git a/tests/xfs/128 b/tests/xfs/128
> > index db5d9a60db..2d2975115e 100755
> > --- a/tests/xfs/128
> > +++ b/tests/xfs/128
> > @@ -34,7 +34,20 @@ margin=160
> >  blksz=65536
> >  real_blksz="$(_get_block_size $testdir)"
> >  blksz_factor=$((blksz / real_blksz))
> > +
> > +# The expected free space numbers in this test require file1 and file4 to share
> > +# the same blocks at the end of the test.  Therefore, we need the allocator to
> > +# give file1 a single extent at the start of the test so that fsr will not be
> > +# tempted to "defragment" a multi-extent file1 or file4.  Defragmenting really
> > +# means rewriting the file, and if that succeeds on either file, we'll have
> > +# unshared the space and there will be too little free space.  Therefore,
> > +# preallocate space to try to produce a single extent.
> > +$XFS_IO_PROG -f -c "falloc 0 $((blks * blksz))" $testdir/file1 >> $seqres.full
> >  _pwrite_byte 0x61 0 $((blks * blksz)) $testdir/file1 >> $seqres.full
> > +sync
> > +
> > +nextents=$($XFS_IO_PROG -c 'stat' $testdir/file1 | grep 'fsxattr.nextents' | awk '{print $3}')
> > +
> >  _cp_reflink $testdir/file1 $testdir/file2
> >  _cp_reflink $testdir/file2 $testdir/file3
> >  _cp_reflink $testdir/file3 $testdir/file4
> > @@ -106,10 +119,23 @@ test $c14 = $c24 || echo "File4 changed by defrag"
> >  
> >  #echo $free_blocks0 $free_blocks1 $free_blocks2 $free_blocks3
> >  
> > -_within_tolerance "free blocks after creating some reflink copies" $free_blocks1 $((free_blocks0 - (blks * blksz_factor) )) $margin -v
> > -_within_tolerance "free blocks after CoW some reflink copies" $free_blocks2 $((free_blocks1 - 2)) $margin -v
> > -_within_tolerance "free blocks after defragging all reflink copies" $free_blocks3 $((free_blocks2 - (blks * 2 * blksz_factor))) $margin -v
> > -_within_tolerance "free blocks after all tests" $free_blocks3 $((free_blocks0 - (blks * 3 * blksz_factor))) $margin -v
> > +freesp_bad=0
> > +
> > +_within_tolerance "free blocks after creating some reflink copies" \
> > +	$free_blocks1 $((free_blocks0 - (blks * blksz_factor) )) $margin -v || freesp_bad=1
> > +
> > +_within_tolerance "free blocks after CoW some reflink copies" \
> > +	$free_blocks2 $((free_blocks1 - 2)) $margin -v || freesp_bad=1
> > +
> > +_within_tolerance "free blocks after defragging all reflink copies" \
> > +	$free_blocks3 $((free_blocks2 - (blks * 2 * blksz_factor))) $margin -v || freesp_bad=1
> > +
> > +_within_tolerance "free blocks after all tests" \
> > +	$free_blocks3 $((free_blocks0 - (blks * 3 * blksz_factor))) $margin -v || freesp_bad=1
> > +
> > +if [ $freesp_bad -ne 0 ] && [ $nextents -gt 0 ]; then

If you decide to commit this, could you change ^^^ this to -gt 1, please?

--D

> > +	echo "free space checks probably failed because file1 nextents was $nextents"
> > +fi
> >  
> >  # success, all done
> >  status=0
> > 
>