[PATCHv6 11/11] iomap: add support for dma aligned direct-io

[Eric Biggers]

On Fri, Jul 22, 2022 at 02:26:05PM -0600, Keith Busch wrote:
> On Fri, Jul 22, 2022 at 06:01:35PM +0000, Eric Biggers wrote:
> > On Fri, Jul 22, 2022 at 08:43:55AM -0600, Keith Busch wrote:
> > > On Fri, Jul 22, 2022 at 12:36:01AM -0700, Eric Biggers wrote:
> > > > [+f2fs list and maintainers]
> > > > 
> > > > On Fri, Jun 10, 2022 at 12:58:30PM -0700, Keith Busch wrote:
> > > > > From: Keith Busch <kbusch at kernel.org>
> > > > > 
> > > > > Use the address alignment requirements from the block_device for direct
> > > > > io instead of requiring addresses be aligned to the block size.
> > > > > 
> > > > > Signed-off-by: Keith Busch <kbusch at kernel.org>
> > > > > Reviewed-by: Christoph Hellwig <hch at lst.de>
> > > > > ---
> > > > >  fs/iomap/direct-io.c | 4 ++--
> > > > >  1 file changed, 2 insertions(+), 2 deletions(-)
> > > > > 
> > > > > diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
> > > > > index 370c3241618a..5d098adba443 100644
> > > > > --- a/fs/iomap/direct-io.c
> > > > > +++ b/fs/iomap/direct-io.c
> > > > > @@ -242,7 +242,6 @@ static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
> > > > >  	struct inode *inode = iter->inode;
> > > > >  	unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
> > > > >  	unsigned int fs_block_size = i_blocksize(inode), pad;
> > > > > -	unsigned int align = iov_iter_alignment(dio->submit.iter);
> > > > >  	loff_t length = iomap_length(iter);
> > > > >  	loff_t pos = iter->pos;
> > > > >  	unsigned int bio_opf;
> > > > > @@ -253,7 +252,8 @@ static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
> > > > >  	size_t copied = 0;
> > > > >  	size_t orig_count;
> > > > >  
> > > > > -	if ((pos | length | align) & ((1 << blkbits) - 1))
> > > > > +	if ((pos | length) & ((1 << blkbits) - 1) ||
> > > > > +	    !bdev_iter_is_aligned(iomap->bdev, dio->submit.iter))
> > > > >  		return -EINVAL;
> > > > >  
> > > > >  	if (iomap->type == IOMAP_UNWRITTEN) {
> > > > 
> > > > I noticed that this patch is going to break the following logic in
> > > > f2fs_should_use_dio() in fs/f2fs/file.c:
> > > > 
> > > > 	/*
> > > > 	 * Direct I/O not aligned to the disk's logical_block_size will be
> > > > 	 * attempted, but will fail with -EINVAL.
> > > > 	 *
> > > > 	 * f2fs additionally requires that direct I/O be aligned to the
> > > > 	 * filesystem block size, which is often a stricter requirement.
> > > > 	 * However, f2fs traditionally falls back to buffered I/O on requests
> > > > 	 * that are logical_block_size-aligned but not fs-block aligned.
> > > > 	 *
> > > > 	 * The below logic implements this behavior.
> > > > 	 */
> > > > 	align = iocb->ki_pos | iov_iter_alignment(iter);
> > > > 	if (!IS_ALIGNED(align, i_blocksize(inode)) &&
> > > > 	    IS_ALIGNED(align, bdev_logical_block_size(inode->i_sb->s_bdev)))
> > > > 		return false;
> > > > 
> > > > 	return true;
> > > > 
> > > > So, f2fs assumes that __iomap_dio_rw() returns an error if the I/O isn't logical
> > > > block aligned.  This patch changes that.  The result is that DIO will sometimes
> > > > proceed in cases where the I/O doesn't have the fs block alignment required by
> > > > f2fs for all DIO.
> > > > 
> > > > Does anyone have any thoughts about what f2fs should be doing here?  I think
> > > > it's weird that f2fs has different behaviors for different degrees of
> > > > misalignment: fail with EINVAL if not logical block aligned, else fallback to
> > > > buffered I/O if not fs block aligned.  I think it should be one convention or
> > > > the other.  Any opinions about which one it should be?
> > > 
> > > It looks like f2fs just falls back to buffered IO for this condition without
> > > reaching the new code in iomap_dio_bio_iter().
> > 
> > No.  It's a bit subtle, so read the code and what I'm saying carefully.  f2fs
> > only supports 4K aligned DIO and normally falls back to buffered I/O; however,
> > for DIO that is *very* misaligned (not even LBS aligned) it returns EINVAL
> > instead.  And it relies on __iomap_dio_rw() returning that EINVAL.
> 
> Okay, I understand the code flow now.
> 
> I tested f2fs direct io with every possible alignment, and it is successful for
> all hardware dma supported alignments and continues to return EINVAL for
> unaligned. Is the concern that it's now returning success in scenarios that
> used to fail? Or is there some other 4k f2fs constraint that I haven't found
> yet?

f2fs has a lot of different options, and edge cases that only occur occasionally
such as data blocks being moved by garbage collection.  So just because
misaligned DIO appears to work doesn't necessarily mean that it actually works.
Maybe the f2fs developers can elaborate on the reason that f2fs always requires
4K alignment for DIO.

- Eric