On Fri, Feb 18, 2022 at 01:55:37PM +0800, Gao Xiang wrote: > Hi Darrick, > > On Thu, Feb 17, 2022 at 09:40:32PM -0800, Darrick J. Wong wrote: > > On Thu, Feb 17, 2022 at 05:55:42PM +0800, Gao Xiang wrote: > > > COW extents are already converted into written real extents after > > > xfs_reflink_convert_cow_locked(), therefore cmap->br_state should > > > reflect it. > > > > > > Otherwise, there is another necessary unwritten convertion > > > triggered in xfs_dio_write_end_io() for direct I/O cases. > > > > > > Signed-off-by: Gao Xiang <hsiangkao@xxxxxxxxxxxxxxxxx> > > > > I /think/ this looks ok. Does it test ok too? AFAICT nothing in the > > iomap/writeback machinery cares the incorrect state because we always > > set IOMAP_F_SHARED (which triggers COW) so I think this is simply a fix > > for directio, like you said? > > Yeah, I think so, from the code logic buffered i/o seems no impacted... > And the unnecessary unwritten convertion under direct i/o takes > noticeable extra overhead in our workloads... > > I checked my last night xfstests, it seems it stops unexpectedly (maybe > due to some environmental problem). I will rerun tests today and > feedback later. On my test environment machine, with linux 5.17-rc4, both w/ and w/o the patch, it fails: generic/594 generic/600 xfs/158 xfs/160 I think no issue with this patch. Thanks, Gao Xiang
