Brian,
Thanks for the review.
You're right. This deviates from the purpose of the test, that, the
original purpose of this patch is to ignore the noise, but this is
more like monitoring the noise. And another point reminding me is this
test is for all file systems, so we can't set it to strict 0 just
because it's all passed on ext.
I have restored the write tolerance back to 5%, and am sending the v3 patch.
Thanks,
Boyang
On Mon, Apr 26, 2021 at 8:08 PM Brian Foster <bfoster@xxxxxxxxxx> wrote:
>
> On Sun, Apr 25, 2021 at 12:53:07PM +0800, Boyang Xue wrote:
> > Hi Brian,
> >
> > Please find my reply below inline.
> >
> > On Fri, Apr 23, 2021 at 7:07 PM Brian Foster <bfoster@xxxxxxxxxx> wrote:
> > >
> > > On Thu, Apr 22, 2021 at 11:31:47PM +0800, bxue@xxxxxxxxxx wrote:
> > > > From: Boyang Xue <bxue@xxxxxxxxxx>
> > > >
> > > > On ext2/ext3, it's expected that several single block metadata reads can occur
> > > > when writing to file in the same cgroup (the stack is like below[1]). The
> > > > purpose of the "write -> read/write" subtest is to make sure the larger pwrite
> > > > is accounted to the correct cgroup, not necessarily enforce that zero bytes are
> > > > read in service of the write. This patch fixes the sub-test in order to tolerate
> > > > small reads in 1st cgroup.
> > > >
> > > > [1] Callchain of the read:
> > > >
> > > > @ext3_read_bio[
> > > > submit_bio+1
> > > > submit_bh_wbc+365
> > > > ext4_read_bh+72
> > > > ext4_get_branch+201
> > > > ext4_ind_map_blocks+382
> > > > ext4_map_blocks+295
> > > > _ext4_get_block+170
> > > > __block_write_begin_int+328
> > > > ext4_write_begin+541
> > > > generic_perform_write+213
> > > > ext4_buffered_write_iter+167
> > > > new_sync_write+345
> > > > vfs_write+438
> > > > __x64_sys_pwrite64+140
> > > > do_syscall_64+51
> > > > entry_SYSCALL_64_after_hwframe+68
> > > > , 5793, 12]: 3
> > > >
> > > > Signed-off-by: Boyang Xue <bxue@xxxxxxxxxx>
> > > > ---
> > > > Hi,
> > > >
> > > > This patch fix the "write -> read/write" sub-test in order to tolerate
> > > > small reads in service of the write (like read metadata).
> > > >
> > > > Change from v1:
> > > > (1) More details in commit log, including example call stack
> > > > (2) Set the fixed tolerance value to 33792 for accuracy
> > > > (3) Update percentage tolerance value to fixed value 0, where doesn't
> > > > fail the test
> > > >
> > > > Tested pass on ext2/ext3/ext4 x 1k/2k/4k blksize.
> > > >
> > > > Thanks,
> > > > Boyang
> > > >
> > > > tests/generic/563 | 20 +++++++++++++-------
> > > > 1 file changed, 13 insertions(+), 7 deletions(-)
> > > >
> > > > diff --git a/tests/generic/563 b/tests/generic/563
> > > > index b113eacf..44394b4b 100755
> > > > --- a/tests/generic/563
> > > > +++ b/tests/generic/563
> > > > @@ -60,6 +60,8 @@ check_cg()
> > > > cgname=$(basename $cgroot)
> > > > expectedread=$2
> > > > expectedwrite=$3
> > > > + readtol=$4
> > > > + writetol=$5
> > > > rbytes=0
> > > > wbytes=0
> > > >
> > > > @@ -71,8 +73,8 @@ check_cg()
> > > > awk -F = '{ print $2 }'`
> > > > fi
> > > >
> > > > - _within_tolerance "read" $rbytes $expectedread 5% -v
> > > > - _within_tolerance "write" $wbytes $expectedwrite 5% -v
> > > > + _within_tolerance "read" $rbytes $expectedread $readtol -v
> > > > + _within_tolerance "write" $wbytes $expectedwrite $writetol -v
> > > > }
> > > >
> > > > # Move current process to another cgroup.
> > > > @@ -113,7 +115,7 @@ $XFS_IO_PROG -c "pread 0 $iosize" -c "pwrite 0 $iosize" -c fsync \
> > > > $SCRATCH_MNT/file >> $seqres.full 2>&1
> > > > switch_cg $cgdir
> > > > $XFS_IO_PROG -c fsync $SCRATCH_MNT/file
> > > > -check_cg $cgdir/$seq-cg $iosize $iosize
> > > > +check_cg $cgdir/$seq-cg $iosize $iosize 5% 5%
> >
> > Here the write tolerance has to be 5% rather than 0, since testing
> > with ext2, the write bytes are slightly larger than $iosize (8392704
> > vs 8388608).
> >
>
> Yeah, for any non-zero values the write tolerance should always be 5%.
> That's the historical behavior of the test.
>
> > > >
> > > > # Write from one cgroup then read and write from a second. Writes are charged to
> > > > # the first group and nothing to the second.
> > > > @@ -126,8 +128,12 @@ $XFS_IO_PROG -c "pread 0 $iosize" -c "pwrite 0 $iosize" $SCRATCH_MNT/file \
> > > > >> $seqres.full 2>&1
> > > > switch_cg $cgdir
> > > > $XFS_IO_PROG -c fsync $SCRATCH_MNT/file
> > > > -check_cg $cgdir/$seq-cg 0 $iosize
> > > > -check_cg $cgdir/$seq-cg-2 0 0
> > > > +# Use a fixed value tolerance for the expected value of zero here
> > > > +# because filesystems might perform a small number of metadata reads to
> > > > +# complete the write. On ext2/3 with 1k block size, the read bytes is
> > > > +# as large as 33792.
> > > > +check_cg $cgdir/$seq-cg 0 $iosize 33792 0
> > >
> > > Shouldn't that last parameter (write tolerance) remain as 5%?
> >
> > In my test on ext2/3/4 with 1k/2k/4k, the write tolerance here always
> > equals $iosize in this sub-test. So make it 0 rather 5%, I hope it
> > will gate future behavior change more sensitively.
> >
>
> I think that deviates from the purpose of the test. I don't think we
> want to filter this change in behavior down through folks' test
> environments just because it happens to be exact in some cases. The
> original purpose of the tolerance was to filter out this kind of noise
> and test the general effectiveness of the cgroup writeback accounting,
> not to capture and document current or future deviations from the exact
> I/O sizes being used.
>
> The problem this patch should fix is that the original test didn't
> properly account for potential metadata I/O to support data I/O by
> virtue of the 5% tolerance being applied to an iosize of 0. Otherwise, I
> don't think it should change the behavior of the test in any way. You
> could always create a new fs-specific test if you wanted to test for
> precise accounting values for specific filesystem(s). Personally, I
> don't think that's really necessary because when such a test fails it
> seems more likely to me we'd end up just having to fix the test over
> changing things on the kernel side, but that's just my .02.
>
> Brian
>
> > >
> > > > +check_cg $cgdir/$seq-cg-2 0 0 0 0
> > > >
> > > > # Read from one cgroup, read & write from a second. Both reads and writes are
> > > > # charged to the first group and nothing to the second.
> > > > @@ -140,8 +146,8 @@ $XFS_IO_PROG -c "pread 0 $iosize" -c "pwrite 0 $iosize" $SCRATCH_MNT/file \
> > > > >> $seqres.full 2>&1
> > > > switch_cg $cgdir
> > > > $XFS_IO_PROG -c fsync $SCRATCH_MNT/file
> > > > -check_cg $cgdir/$seq-cg $iosize $iosize
> > > > -check_cg $cgdir/$seq-cg-2 0 0
> > > > +check_cg $cgdir/$seq-cg $iosize $iosize 5% 0
> > >
> > > And here too? Otherwise the patch LGTM.
> >
> > The same reason as above.
> >
> > If the write tolerance mismatch confuses future readers, I can put a
> > comment to explain. Maybe a single comment like this?
> >
> > ```
> > # Read and write from a single group.
> > echo "read/write"
> > reset
> > switch_cg $cgdir/$seq-cg
> > $XFS_IO_PROG -c "pread 0 $iosize" -c "pwrite 0 $iosize" -c fsync \
> > $SCRATCH_MNT/file >> $seqres.full 2>&1
> > switch_cg $cgdir
> > $XFS_IO_PROG -c fsync $SCRATCH_MNT/file
> > # write bytes is slightly larger than $iosize on ext2 <=== new comment
> > check_cg $cgdir/$seq-cg $iosize $iosize 5% 5%
> > ```
> >
> > Thanks for the review!
> >
> > -Boyang
> >
> > > Brian
> > >
> > > > +check_cg $cgdir/$seq-cg-2 0 0 0 0
> > > >
> > > > echo "-io" > $cgdir/cgroup.subtree_control || _fail "subtree control"
> > > >
> > > > --
> > > > 2.27.0
> > > >
> > >
> >
>
