On Thu, Nov 06, 2014 at 07:49:08AM -0500, Brian Foster wrote:
> There's a simple test for that condition, as noted in my previous mail
> as well in case you missed it. Again, that probably calls out that we
> could be doing better unit testing of bulkstat in xfstests. At the very
> least we probably need some bulkstat inode count validation against a
> known data set.
That's exactly what I've been running to find this latest problem.
But a 500TB filesystem with 10 million inodes in it is a bit beyond
xfstests. ANd that only showed up the problem in 4 AGs out of 500,
so with smaller filesystems there's a good chance that this would
have also been missed....
> xfsdump testing is obviously important, but if bulkstat
> is broken then we clearly can't expect xfsdump to work (and debugging
> the former via the latter appears to be quite painful).
We have a bulkstat command in xfstests. And it can be used to
comapre the output against a stat of the file. But it can't detect
missing inodes and I don't think we can start at arbitrary inodes,
either. So it needs work to be able to be used in unit tests.