On Sun, Jul 26, 2020 at 09:47:51AM +1000, Dave Chinner wrote: > On Fri, Jul 24, 2020 at 10:41:32AM -0700, Eric Biggers wrote: > > But again, as far as I can tell, fs/iomap/direct-io.c currently *does* guarantee > > that *if* the input is fully filesystem-block-aligned and if blocksize <= > > PAGE_SIZE, then the issued I/O is also filesystem-block-aligned. > > Please listen to what I'm saying, Eric. > > The -current iomap implementation- may provide that behaviour. That > doesn't mean we guarantee that behaviour. i.e. the iomap -design- > does not guaranteee that behaviour, and we don't guarantee such > behaviour into the future. And we won't guarantee this behaviour - > even though the current implementation may provide it - because the > rest of the IO stack below iomap does not provide iomap with that > guarantee. > > Hence if iomap cannot get a guarantee that IO it issues won't get > split at some arbitrary boundary, it cannot provide filesystems with > that guarantee. BTW, if you want iomap_dio_rw() to provide an arbitrary bio alignment guarantee at the iomap layer, then it should be returned in the iomap along with the extent mapping. That could then be used instead of the bdev logical block size. That won't guarantee the behaviour of the rest of the stack, but it would provide a defined IO submission behaviour that iomap would have to guarantee into the future... That would also remove the need to duplicate the alignment checks in the filesystem for fscrypt DIO... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx
