fix a DEBUG-only assert failure in xfs/538

[Christoph Hellwig <hch@lst.de>]

Hi all,

when testing with very small rtgroups I've seen relatively frequent
failures in xfs/538 where an assert about the da block type triggers
that should be entirely impossible to trigger by the expected code
flow.

It turns out for this two things had to come together:  a bug in the
attr code to uses ENOSPC to signal a condition that is not related
to run out free blocks, but which can also be triggered when we
actually run out of free blocks, and a debug in the DEBUG only
xfs_bmap_exact_minlen_extent_alloc allocator trigger only by the
specific error injection used in this and a few other tests.

This series tries to fix both issues and clean up the surrounding
code a bit to make it more obvious.

Diffstat;
 xfs_attr.c      |  178 ++++++++++++++++++++++----------------------------------
 xfs_attr_leaf.c |   40 ++++++------
 xfs_attr_leaf.h |    2 
 xfs_bmap.c      |  134 +++++++++++++-----------------------------
 xfs_da_btree.c  |    5 -
 5 files changed, 143 insertions(+), 216 deletions(-)