On Fri, Jul 22, 2022 at 12:12:23AM -0700, Eric Biggers wrote: > -static bool ext4_dio_supported(struct kiocb *iocb, struct iov_iter *iter) > +/* > + * Returns %true if the given DIO request should be attempted with DIO, or > + * %false if it should fall back to buffered I/O. > + * > + * DIO isn't well specified; when it's unsupported (either due to the request > + * being misaligned, or due to the file not supporting DIO at all), filesystems > + * either fall back to buffered I/O or return EINVAL. For files that don't use > + * any special features like encryption or verity, ext4 has traditionally > + * returned EINVAL for misaligned DIO. iomap_dio_rw() uses this convention too. > + * In this case, we should attempt the DIO, *not* fall back to buffered I/O. > + * > + * In contrast, in cases where DIO is unsupported due to ext4 features, ext4 > + * traditionally falls back to buffered I/O. > + * > + * This function implements the traditional ext4 behavior in all these cases. Heh. I had been under the impression that misaligned I/O fell back to buffered I/O for ext4, since that's what a lot of historical Unix systems did. Obviously, it's not something I've tested since "you should never do that". There's actually some interesting discussion about what Linux *should* be doing in the futre in this discussion: https://patchwork.ozlabs.org/project/linux-ext4/patch/1461472078-20104-1-git-send-email-tytso@xxxxxxx/ Including the following from Christoph Hellwig: https://patchwork.ozlabs.org/project/linux-ext4/patch/1461472078-20104-1-git-send-email-tytso@xxxxxxx/#1335016 > I've been doing an audit of our direct I/O implementations, and most > of them does some form of transparent fallback, including some that > only pretend to support O_DIRECT, but do anything special for it at all, > while at the same time we go through greast efforts to check a file > system actualy supports direct I/O, leading to nasty no-op ->direct_IO > implementations as we even got that abstraction wrong. > > At this point I wonder if we should simply treat O_DIRECT as a hint > and always allow it, and just let the file system optimize for it > (skip buffering, require alignment, relaxed Posix atomicy requirements) > if it is set. The thread also mentioned XFS_IOC_DIOINFO and how We Really Should have something with equivalent functionality to the VFS --- six years ago. :-) Anyway, this change to ext4 looks good. Acked-by: Theodore Ts'o <tytso@xxxxxxx> - Ted
