fix stale delalloc punching for COW I/O v3

[Christoph Hellwig <hch@lst.de>]

Hi all,

this is another fallout from the zoned XFS work, which stresses the XFS
COW I/O path very heavily.  It affects normal I/O to reflinked files as
well, but is very hard to hit there.

The main problem here is that we only punch out delalloc reservations
from the data fork, but COW I/O places delalloc extents into the COW
fork, which means that it won't get punched out forshort writes.

Changes since v2:
 - drop the patches already merged and rebased to latest Linus' tree
 - moved taking invalidate_lock from iomap to the caller to avoid a
   too complicated locking protocol
 - better document the xfs_file_write_zero_eof return value
 - fix a commit log typo

Changes since v1:
 - move the already reviewed iomap prep changes to the beginning in case
   Christian wants to take them ASAP
 - take the invalidate_lock for post-EOF zeroing so that we have a
   consistent locking pattern for zeroing.

Diffstat:
 Documentation/filesystems/iomap/operations.rst |    2 
 fs/iomap/buffered-io.c                         |  111 ++++++-------------
 fs/xfs/xfs_aops.c                              |    4 
 fs/xfs/xfs_bmap_util.c                         |   10 +
 fs/xfs/xfs_bmap_util.h                         |    2 
 fs/xfs/xfs_file.c                              |  146 +++++++++++++++----------
 fs/xfs/xfs_iomap.c                             |   65 +++++++----
 include/linux/iomap.h                          |   20 ++-
 8 files changed, 196 insertions(+), 164 deletions(-)