A couple of fixes identified for xfs_repair follow. The first is a
regression introduced by the discontiguous buffer support in the
prefetch code, which Eric Sandeen found. I've written a slightly
different fix for it than Eric did, but the concept for the fix is
all Eric's work.
The second patch is not actually a regression, but the filesystem
that triggered the above regression can't actually be repaired in a
sane amount of time because of an array insertion algorithm with
exponential cost. It's trivial to fix, and it means xfs_repair can
actually run to completion quickly rather than burn CPU for hours
getting nothing done.
These fixes mean I'll probably release a 3.2.0-rc3 once they've been
reviewed and go into the tree....