Re: [PATCH v2] nfs: test files written size as expected

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

On Thu, Aug 03, 2023 at 01:24:18AM +0800, Zorro Lang wrote:
> On Wed, Aug 02, 2023 at 09:36:40AM -0700, Darrick J. Wong wrote:
> > On Wed, Aug 02, 2023 at 01:46:46PM +0800, Zorro Lang wrote:
> > > Test nfs and its underlying fs, make sure file size as expected
> > > after writting a file, and the speculative allocation space can
> > > be shrunken.
> > > 
> > > Signed-off-by: Zorro Lang <zlang@kernel.org>
> > > ---
> > > 
> > > Last year I sent a patch to fstests@, but it sometimes fails on the upstream
> > > kernel that year:
> > > 
> > >   https://lore.kernel.org/fstests/Y3vTbHqT64gsQ573@magnolia/
> > > 
> > > And we didn't get a proper reason for that, so that patch was blocked. Now
> > > I found this case test passed on current upstream linux [1] (after loop
> > > running it a whole night). So I think it's time to rebase and re-send this
> > > patch to get review.
> > > 
> > > Thanks,
> > > Zorro
> > > 
> > > [1]
> > > FSTYP         -- nfs
> > > PLATFORM      -- Linux/x86_64 xxxx 6.5.0-rc4 #1 SMP PREEMPT_DYNAMIC Tue Aug  1 15:32:55 EDT 2023
> > > MKFS_OPTIONS  -- xxxx.redhat.com:/mnt/xfstests/scratch/nfs-server
> > > MOUNT_OPTIONS -- -o vers=4.2 -o context=system_u:object_r:root_t:s0 xxxx.redhat.com:/mnt/xfstests/scratch/nfs-server /mnt/xfstests/scratch/nfs-client
> > > 
> > > nfs/002 4s ...  4s
> > > Ran: nfs/002
> > > Passed all 1 tests
> > > 
> > >  tests/nfs/002     | 46 ++++++++++++++++++++++++++++++++++++++++++++++
> > >  tests/nfs/002.out |  2 ++
> > >  2 files changed, 48 insertions(+)
> > >  create mode 100755 tests/nfs/002
> > >  create mode 100644 tests/nfs/002.out
> > > 
> > > diff --git a/tests/nfs/002 b/tests/nfs/002
> > > new file mode 100755
> > > index 00000000..b4b6554c
> > > --- /dev/null
> > > +++ b/tests/nfs/002
> > > @@ -0,0 +1,46 @@
> > > +#! /bin/bash
> > > +# SPDX-License-Identifier: GPL-2.0
> > > +# Copyright (c) 2023 Red Hat, Inc.  All Rights Reserved.
> > > +#
> > > +# FS QA Test 002
> > > +#
> > > +# Make sure nfs gets expected file size after writting a big sized file. It's
> > > +# not only testing nfs, test its underlying fs too. For example a known old bug
> > > +# on xfs (underlying fs) caused nfs get larger file size (e.g. 16M) after
> > > +# writting 10M data to a file. It's fixed by a series of patches around
> > > +# 579b62faa5fb16 ("xfs: add background scanning to clear eofblocks inodes")
> > 
> > Er... has this been banging around in the trunk for 11 years? ;)
> 
> Yeah, that's an old enough test case :-D I tried to tidy our internal test cases,
> felt this case can be in fstests.
> 
> > 
> > > +#
> > > +. ./common/preamble
> > > +_begin_fstest auto quick rw
> > > +
> > > +# real QA test starts here
> > > +_supported_fs nfs
> > > +# Need a series of patches related with this patch
> > > +_fixed_by_kernel_commit 579b62faa5fb16 \
> > > +	"xfs: add background scanning to clear eofblocks inodes"
> > > +_require_test
> > > +
> > > +localfile=$TEST_DIR/testfile.$seq
> > > +rm -rf $localfile
> > > +
> > > +$XFS_IO_PROG -f -t -c "pwrite 0 10m" -c "fsync" $localfile >>$seqres.full 2>&1
> > > +block_size=`stat -c '%B' $localfile`
> > > +iblocks_expected=$((10 * 1024 * 1024 / $block_size))
> > > +# Try several times for the speculative allocated file size can be shrunken
> > > +res=1
> > > +for ((i=0; i<10; i++));do
> > > +	iblocks_real=`stat -c '%b' $localfile`
> > > +	if [ "$iblocks_expected" = "$iblocks_real" ];then
> > 
> > What happens if real < expected?  Should there be some sort of bail out
> > for unexpected things like that?
> 
> Hmm... I never thought that. I saw the real >= expected, is there any
> chance to get real < expected?

<shrug> Suppose the NFS server is running on top of a filesystem that
supports compression and i_blocks as returned by stat reflects that?

--D

> > 
> > > +		res=0
> > > +		break
> > > +	fi
> > > +	sleep 10
> > > +done
> > 
> > Though I guess the runtime is capped at ~100s so maybe it doesn't
> > matter practically.
> 
> Mostly the test done in several seconds in my testing:
> 
> FSTYP         -- nfs
> PLATFORM      -- Linux/x86_64 hp-dl360g9-06 6.5.0-rc4 #1 SMP PREEMPT_DYNAMIC Tue Aug  1 15:32:55 EDT 2023
> MKFS_OPTIONS  -- hp-dl360g9-06.rhts.eng.pek2.redhat.com:/mnt/xfstests/scratch/nfs-server
> MOUNT_OPTIONS -- -o vers=4.2 -o context=system_u:object_r:root_t:s0 hp-dl360g9-06.rhts.eng.pek2.redhat.com:/mnt/xfstests/scratch/nfs-server /mnt/xfstests/scratch/nfs-client
> 
> nfs/002 5s ...  4s
> Ran: nfs/002
> Passed all 1 tests

Doesn't xfs remove the speculative preallocations every time a write fd
is closed?

Yes, it does do that:
https://lore.kernel.org/linux-xfs/155259894034.30230.7188877605950498518.stgit@magnolia/

IOWs, how is this test actually checking the behavior of background
blockgc clearing out speculative preallocations?

> > (What happens if xfs blockgc only runs every 5 minutes?)
> 
> How can can make that happen? If the 100s isn't enough, is there an upper
> limit, or how to make an upper limit?

There's no way to tell over NFS...

--D

> 
> Thanks,
> Zorro
> 
> > 
> > --D
> > 
> > > +if [ $res -ne 0 ];then
> > > +	echo "Write $iblocks_expected blocks, but get $iblocks_real blocks"
> > > +fi
> > > +
> > > +echo "Silence is golden"
> > > +# success, all done
> > > +status=0
> > > +exit
> > > diff --git a/tests/nfs/002.out b/tests/nfs/002.out
> > > new file mode 100644
> > > index 00000000..61705c7c
> > > --- /dev/null
> > > +++ b/tests/nfs/002.out
> > > @@ -0,0 +1,2 @@
> > > +QA output created by 002
> > > +Silence is golden
> > > -- 
> > > 2.40.1
> > > 
> > 
>