On Wed, Aug 30, 2017 at 7:44 AM, Dave Chinner [off-list ref] wrote:

On Wed, Aug 30, 2017 at 06:16:35AM +0200, Luis R. Rodriguez wrote:

quoted

However this still begs the question of what tests we should run
prior and after the full set, and if we had any missing test what
we should add, or if we've considered that.

We have nothing that actually tests that the on-disk filesystem
format values match what is specified on the command line, so
testing/validation has all been completely manual. I'm not sure we
can automate it, given the math needed to check that all the fields
in the superblock are correct and sane for arbitrary mkfs
configurations on arbitrary block devices.

In the end I basically ran some configs with an old mkfs, dumped the
superblock with xfs_db for each config, then did the same with the
new mkfs. I then diffed the output to see if anything came out
differently. I fixed all the things I found that were different.  I
ran the checks on 4k and 512 byte sector devices, just to make sure
nothing went wrong with using probed device sector sizes by default.
Then I diffed logprint output for different log options (e.g. sector
size, lsunit, etc) to make sure everything was correct, etc. I
mounted each filesystem, ran repair on them, etc so that both the
kernel and repair code could validate the format, too (the kernel
found the first log formatting bug via CRC errors).  And, of course,
I ran xfstests over a variety of configs - different block sizes,
rmap, DAX, etc.

Maybe you can come up with a way of automating this, but for a
one-off piece of work that affects a point-in-time snapshot of mkfs
functionality, I'm not sure it's worth the effort to try to make a
generic test to do this sort of thing.

I actually tried to automate this kind of testing. Something similar
to xfs/191, but with testing if the physically written fs is valid and
expected. But I didn't find a way that would work with xfstests
without hacking it to use some other version of mkfs.xfs manually, and
without doing all the math craziness that would be so error-prone...

Anyway, I'm on the sea until 10th September. I might be able to
continue my review and write something tomorrow (Friday) in the
morning before I depart, but more likely I will keep all my comments
until after I return, and see what changes in the meantime. But so
far, it looks like a good job, Dave. :-)

Cheers,
Jan