On Wed, Jul 08, 2020 at 04:51:27PM +1000, Dave Chinner wrote: > On Tue, Jul 07, 2020 at 03:00:30PM +0200, Christoph Hellwig wrote: > > On Tue, Jul 07, 2020 at 01:57:05PM +0100, Matthew Wilcox wrote: > > > Indeed, I'm in favour of not invalidating > > > the page cache at all for direct I/O. For reads, I think the page cache > > > should be used to satisfy any portion of the read which is currently > > > cached. For writes, I think we should write into the page cache pages > > > which currently exist, and then force those pages to be written back, > > > but left in cache. > > > > Something like that, yes. > > So are we really willing to take the performance regression that > occurs from copying out of the page cache consuming lots more CPU > than an actual direct IO read? Or that direct IO writes suddenly > serialise because there are page cache pages and now we have to do > buffered IO? > > Direct IO should be a deterministic, zero-copy IO path to/from > storage. Using the CPU to copy data during direct IO is the complete > opposite of the intended functionality, not to mention the behaviour > that many applications have been careful designed and tuned for. Direct I/O isn't deterministic though. If the file isn't shared, then it works great, but as soon as you get mixed buffered and direct I/O, everything is already terrible. Direct I/Os perform pagecache lookups already, but instead of using the data that we found in the cache, we (if it's dirty) write it back, wait for the write to complete, remove the page from the pagecache and then perform another I/O to get the data that we just wrote out! And then the app that's using buffered I/O has to read it back in again. Nobody's proposing changing Direct I/O to exclusively work through the pagecache. The proposal is to behave less weirdly when there's already data in the pagecache. I have had an objection raised off-list. In a scenario with a block device shared between two systems and an application which does direct I/O, everything is normally fine. If one of the systems uses tar to back up the contents of the block device then the application on that system will no longer see the writes from the other system because there's nothing to invalidate the pagecache on the first system. Unfortunately, this is in direct conflict with the performance problem caused by some little arsewipe deciding to do: $ while true; do dd if=/lib/x86_64-linux-gnu/libc-2.30.so iflag=direct of=/dev/null; done ... doesn't hurt me because my root filesystem is on ext4 which doesn't purge the cache. But anything using iomap gets all the pages for libc kicked out of the cache, and that's a lot of fun.
