After XFS switching to iomap based DIO (commit acdda3aa ("xfs:
use iomap_dio_rw")), I started to notice dio29/dio30 tests failures
from LTP run on ppc64 hosts, and they can be reproduced on x86_64
hosts with 512B/1k block size XFS too.

dio29	diotest3 -b 65536 -n 100 -i 1000 -o 1024000
dio30	diotest6 -b 65536 -n 100 -i 1000 -o 1024000

The failure message is like:
bufcmp: offset 0: Expected: 0x62, got 0x0
diotest03    1  TPASS  :  Read with Direct IO, Write without
diotest03    2  TFAIL  :  diotest3.c:142: comparsion failed; child=98 offset=1425408
diotest03    3  TFAIL  :  diotest3.c:194: Write Direct-child 98 failed

Direct write wrote 0x62 but buffer read got zero. This is because,
when doing direct write to a hole or preallocated file, we
invalidate the page caches before converting the extent from
unwritten state to normal state, which is done by
iomap_dio_complete(), thus leave a window for other buffer reader to
cache the unwritten state extent.

Consider this case, with sub-page blocksize XFS, two processes are
direct writing to different blocksize-aligned regions (say 512B) of
the same preallocated file, and reading the region back via buffered
I/O to compare contents.

process A, region [0,512]		process B, region [512,1024]
xfs_file_write_iter
 xfs_file_aio_dio_write
  iomap_dio_rw
   iomap_apply
   invalidate_inode_pages2_range
   					xfs_file_write_iter
				 	xfs_file_aio_dio_write
					  iomap_dio_rw
					   iomap_apply
					   invalidate_inode_pages2_range
					   iomap_dio_complete
					xfs_file_read_iter
					 xfs_file_buffered_aio_read
					  generic_file_read_iter
					   do_generic_file_read
					    <readahead fills pagecache with 0>
   iomap_dio_complete
xfs_file_read_iter
 <read gets 0 from pagecache>

Process A first invalidates page caches, at this point the
underlying extent is still in unwritten state (iomap_dio_complete
not called yet), and process B finishs direct write and populates
page caches via readahead, which caches zeros in page for region A,
then process A reads zeros from page cache, instead of the actual
data.

Fix it by invalidating page caches after converting unwritten extent
to make sure we read content from disk after extent state changed,
as what we did before switching to iomap based dio.

Also introduce a new 'start' variable to save the original write
offset (iomap_dio_complete() updates iocb->ki_pos), and a 'err'
variable for invalidating caches result, cause we can't reuse 'ret'
anymore.
Signed-off-by: Eryu Guan <eguan@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>