On Tue, Oct 27, 2020 at 06:07:31PM +0000, Christoph Hellwig wrote: > On Mon, Oct 19, 2020 at 12:55:01PM -0400, Brian Foster wrote: > > On Thu, Oct 15, 2020 at 10:47:00AM +0100, Christoph Hellwig wrote: > > > I don't think we can solve this properly. Due to the racyness we can > > > always err one side. The beauty of treating all the uptodate pages > > > as present data is that we err on the safe side, as applications > > > expect holes to never have data, while "data" could always be zeroed. > > > > > > > I don't think that's quite accurate. Nothing prevents a dirty page from > > being written back and reclaimed between acquiring the (unwritten) > > mapping and doing the pagecache scan, so it's possible to present valid > > data (written to the kernel prior to a seek) as a hole with the current > > code. > > True. I guess we need to go back and do another lookup to fully > solve this problem. That doesn't change my opinion that this patch > makes the problem worse. > Yeah. I think it's possible to at least have some internal consistency (i.e. while we're under locks) if we check the page state first or somehow or another jump back out of the iomap_apply() sequence to do so. I hadn't thought about it a ton since the goal of these patches was to address the post-eof zeroing problem vs. fix seek data/hole. Brian
