Re: [PATCH 5/5] update i_disksize coherently with block allocation

[Dmitry Monakhov <dmonakhov@openvz.org>]

On Sat, 23 Aug 2014 18:00:29 -0400, "Theodore Ts'o" <tytso@mit.edu> wrote:
> On Fri, Aug 22, 2014 at 03:32:27PM +0400, Dmitry Monakhov wrote:
> > Writeback call trace looks like follows:
> > ext4_writepages
> >  while(nr_pages)
> >  ->journal_start
> >  ->mpage_map_and_submit_extent -> may alloc some blocks
> >    ->mpage_map_one_extent
> >  ->journal_stop
> > In case of delalloc block i_disksize may be less than i_size. So we have to
> > update i_disksize each time we allocated and submitted some blocks beyond
> > i_disksize. And we MUST update it in the same transaction, otherwise this
> > result in fs-inconsistency in case of upcoming power-failure.
> > 
> > Another possible way to fix that issue is to insert inode to orhphan list
> > on ext4_writepages entrance.
> > 
> > testcase: xfstest generic/019
> > 
> > Signed-off-by: Dmitry Monakhov <dmonakhov@openvz.org>
> 
> Hi Dmitry, were you seeing generic/019 fail before this patch series?
> I've been trying to build a kernel with CONFIG_FAIL_MAKE_REQUEST and I
> haven't been able to get generic/019 to fail on me.  Is there
> something else we need in order to reliably trigger the test fail?
As usual this kind of test are not 100% reliable, I've saw failures from
time to time. But I've assumed that it was side effect of incorrect
error detection in e2fsck introduced d3f32c2db8f11, But this week i've
rechecked e2fsck and found that condition was fixed and it is correct.
In order to speedup testing I use ram dev:
options brd rd_nr=4 rd_size=10485760 part_show=1
TEST_DEV=/dev/ram0
SCRATCH_DEV=/dev/ram1
And run several rounds for this test:
for ((i=0;i<20;i++));do ./check generic/019 || break ;done

You also can increase probability by playing with fsstress options
--- a/tests/generic/019
+++ b/tests/generic/019
@@ -135,7 +135,7 @@ FSSTRESS_AVOID="$FSSTRESS_AVOID -ffsync=0 -fsync=0
-ffdatasync=0 -f setattr=1"
 _workout()
 {
        out=$SCRATCH_MNT/fsstress.$$
-       args=`_scale_fsstress_args -p 1 -n999999999 -f setattr=0
$FSSTRESS_AVOID -d $out`
+       args=`_scale_fsstress_args -p 8 -n999999999 -f setattr=0
$FSSTRESS_AVOID -d $out`
        echo ""
        echo "Start fsstress.."
        echo ""

And finally the cherry on top of this cake I've found that this test
provoke orphan list corruption or dangling inodes after failure.
fsck 1.43-WIP (09-Jul-2014)
e2fsck 1.43-WIP (09-Jul-2014)
Pass 1: Checking inodes, blocks, and sizes
Deleted inode 43792 has zero dtime.  Fix<y>? no
Inodes that were part of a corrupted orphan linked list found.  Fix<y>?
no
Inode 493817 was part of the orphaned inode list.  IGNORED.
Pass 2: Checking directory structure
Pass 3: Checking directory connectivity
Pass 4: Checking reference counts
Pass 5: Checking group summary information
Block bitmap differences:  -148712 -148714
Fix<y>? no
Inode bitmap differences:  -43792 -493817
Fix<y>? no

/dev/ram1: ********** WARNING: Filesystem still has errors **********

/dev/ram1: 201244/655360 files (0.0% non-contiguous), 409632/10485760
blocks
[root@ts105 xfstests-dev.git2]# INO=493817
[root@ts105 xfstests-dev.git2]# debugfs /dev/ram1 -R "ex <$INO>" ; \
            debugfs /dev/ram1 -R "stat <$INO>" ; debugfs /dev/ram1 -R "ncheck $INO"
debugfs 1.43-WIP (09-Jul-2014)
Level Entries       Logical            Physical Length Flags
 0/ 0   1/  1     0 -     0   148712 -   148712      1 
debugfs 1.43-WIP (09-Jul-2014)
Inode: 493817   Type: symlink    Mode:  0777   Flags: 0x80000
Generation: 4038911591    Version: 0x00000000:00000001
User:     0   Group:     0   Size: 638
File ACL: 0    Directory ACL: 0
Links: 0   Blockcount: 2
Fragment:  Address: 0    Number: 0    Size: 0
 ctime: 0x53fae8e8:ea861dc0 -- Mon Aug 25 11:42:32 2014
 atime: 0x53fae8e8:ea861dc0 -- Mon Aug 25 11:42:32 2014
 mtime: 0x53fae8e8:ea861dc0 -- Mon Aug 25 11:42:32 2014
crtime: 0x53fae8e8:ea861dc0 -- Mon Aug 25 11:42:32 2014
dtime: 0x0000ab10 -- Thu Jan  1 15:09:52 1970
Size of extra inode fields: 28
EXTENTS:
(0):148712
debugfs 1.43-WIP (09-Jul-2014)
Inode   Pathname

I saw this effect with different file types (synmlink,chdev,regfile)
>From my findings we lost newly created inode during creation.
Actually code is very simple, but at this moment I can not find why and
where this happen.