Re: [RFC] iomap: fix race between readahead and direct write

[Darrick J. Wong]

On Fri, Jan 17, 2020 at 12:05:36PM +0100, Jan Kara wrote:
> On Fri 17-01-20 17:39:03, yukuai (C) wrote:
> > On 2020/1/16 23:32, Jan Kara wrote:
> > > Thanks for the report and the patch. But the data integrity when mixing
> > > buffered and direct IO like this is best effort only. We definitely do not
> > > want to sacrifice performance of common cases or code complexity to make
> > > cases like this work reliably.
> > 
> > In the patch, the only thing that is diffrent is that iomap_begin() will
> > be called for each page. However, it seems the performance in sequential
> > read didn't get worse. Is there a specific case that the performance
> > will get worse?
> 
> Well, one of the big points of iomap infrastructure is that you call
> filesystem once to give you large extent instead of calling it to provide
> allocation for each page separately. The additional CPU overhead will be
> visible if you push the machine hard enough. So IMHO the overhead just is
> not worth it for a corner-case like you presented. But that's just my
> opinion, Darrick and Christoph are definitive arbiters here...

Does the problem go away if you apply[1]?  If I understand the race
correctly, marking the extents unwritten and leaving them that way until
after we've written the disk should eliminate the exposure vector...? :)

[1] https://lore.kernel.org/linux-xfs/157915535059.2406747.264640456606868955.stgit@magnolia/

--D

> 								Honza
> 
> -- 
> Jan Kara <jack@xxxxxxxx>
> SUSE Labs, CR