On Tue, Feb 28, 2017 at 08:51:39AM -0800, Darrick J. Wong wrote: > Hm... is fio_alignment is specified in units of bytes? Yes. > If so, then > shouldn't this be a __u32 so that we can handle some weird future device > that wants, say, 1MB alignment for its atomic IO? That would be pretty useless. Anything bigger than sector / block size would not really be usable for typical applications. > Though, now that I look at the XFS ioinfo patch, I guess fio_alignment > is set only for O_DIRECT files? Yes. > So it's really the required alignment > for directio operations. For buffered I/O we can write at byte granularity and still use the atomic commits, but for direct I/O we can only COW at block size granularity.
