On Thu, Sep 21, 2023 at 8:39 PM Naohiro Aota <naohiro.aota@xxxxxxx> wrote: > > Running btrfs/239 on a zoned device often fails with the following error. > > btrfs/239 5s ... - output mismatch (see /host/btrfs/239.out.bad) > --- tests/btrfs/239.out 2023-09-21 16:56:37.735204924 +0900 > +++ /host/btrfs/239.out.bad 2023-09-21 18:22:45.401433408 +0900 > @@ -1,4 +1,6 @@ > QA output created by 239 > +/testdir/dira still exists > +/dira does not exists > File SCRATCH_MNT/testdir/file1 data: > 0000000 ab ab ab ab ab ab ab ab ab ab ab ab ab ab ab ab > * > ... > > This happens because "testdir" and "dira" are not logged on the first fsync > (fsync $SCRATCH_MNT/testdir), but are written as a full commit. That > prevents updating the log on "mv" time, leaving them pre-mv state. > > The full commit is induced by the creation of a new block group. On the > zoned mode, we use a dedicated block group for tree-log. That block group > is created on-demand or assigned to a metadata block group if there is > none. On the first mount of a file system, we need to create one because > there is only one metadata block group available for the regular > metadata. That creation of a new block group forces tree-log to be a full > commit on that transaction, which prevents logging "testdir" and "dira". > > Fix the issue by calling fsync before the first "sync", which creates the > dedicated block group and let the files be properly logged. > > Signed-off-by: Naohiro Aota <naohiro.aota@xxxxxxx> Reviewed-by: Filipe Manana <fdmanana@xxxxxxxx> Looks reasonable, thanks. > --- > tests/btrfs/239 | 12 ++++++++++++ > 1 file changed, 12 insertions(+) > > diff --git a/tests/btrfs/239 b/tests/btrfs/239 > index 3fbeaedd2c39..5a2dbe58d5fb 100755 > --- a/tests/btrfs/239 > +++ b/tests/btrfs/239 > @@ -83,6 +83,18 @@ done > # > mkdir $SCRATCH_MNT/testdir/dira > > +# This fsync is for the zoned mode. On the zoned mode, we use a dedicated block > +# group for tree-log. That block group is created on-demand or assigned to a > +# metadata block group if there is none. On the first mount of a file system, we > +# need to create one because there is only one metadata block group available > +# for the regular metadata. That creation of a new block group forces tree-log > +# to be a full commit on that transaction, which prevents logging "testdir" and > +# "dira" and screws up the result. > +# > +# Calling fsync here will create the dedicated block group, and let them be > +# logged. > +$XFS_IO_PROG -c "fsync" $SCRATCH_MNT > + > # Make sure everything done so far is durably persisted. > sync > > -- > 2.42.0 >
