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

[Eric Biggers]

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.

Relying on __iomap_dio_rw() in that way is definitely a bad design on f2fs's
part (and I messed that up when switching f2fs from fs/direct-io.c to iomap).
The obvious fix is to just have f2fs do the LBS alignment check itself.

But I think that f2fs shouldn't have different behavior for different levels of
misalignment in the first place, so I was wondering if anyone had any thoughts
on which behavior (EINVAL or fallback to buffered I/O) should be standardized on
in all cases, at least for f2fs.  There was some discussion about this sort of
thing for ext4 several years ago on the thread
https://lore.kernel.org/linux-ext4/1461472078-20104-1-git-send-email-tytso@mit.edu/T/#u,
but it didn't really reach a conclusion.  I'm wondering if the f2fs maintainers
have any thoughts about why the f2fs behavior is as it is.  I.e. is it just
accidental, or are there specific reasons...

- Eric