xfs: remove the i_new_size field in struct xfs_inode
authorChristoph Hellwig <hch@infradead.org>
Sun, 18 Dec 2011 20:00:12 +0000 (20:00 +0000)
committerBen Myers <bpm@sgi.com>
Tue, 17 Jan 2012 21:10:19 +0000 (15:10 -0600)
commit2813d682e8e6a278f94817429afd46b30875bb6e
treed865b04ec89076b692a922b7f5fced9be0458f47
parentce7ae151ddada3dbf67301464343c154903166b3
xfs: remove the i_new_size field in struct xfs_inode

Now that we use the VFS i_size field throughout XFS there is no need for the
i_new_size field any more given that the VFS i_size field gets updated
in ->write_end before unlocking the page, and thus is always uptodate when
writeback could see a page.  Removing i_new_size also has the advantage that
we will never have to trim back di_size during a failed buffered write,
given that it never gets updated past i_size.

Note that currently the generic direct I/O code only updates i_size after
calling our end_io handler, which requires a small workaround to make
sure di_size actually makes it to disk.  I hope to fix this properly in
the generic code.

A downside is that we lose the support for parallel non-overlapping O_DIRECT
appending writes that recently was added.  I don't think keeping the complex
and fragile i_new_size infrastructure for this is a good tradeoff - if we
really care about parallel appending writers we should investigate turning
the iolock into a range lock, which would also allow for parallel
non-overlapping buffered writers.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
fs/xfs/xfs_aops.c
fs/xfs/xfs_file.c
fs/xfs/xfs_iget.c
fs/xfs/xfs_inode.h
fs/xfs/xfs_trace.h