[PATCH RFC 11/16] fs: iomap: Atomic write support

Dave Chinner david at fromorbit.com
Wed May 3 22:00:06 PDT 2023


On Wed, May 03, 2023 at 06:38:16PM +0000, John Garry wrote:
> Add support to create bio's whose bi_sector and bi_size are aligned to and
> multiple of atomic_write_unit, respectively.
> 
> When we call iomap_dio_bio_iter() -> bio_iov_iter_get_pages() ->
> __bio_iov_iter_get_pages(), we trim the bio to a multiple of
> atomic_write_unit.
> 
> As such, we expect the iomi start and length to have same size and
> alignment requirements per iomap_dio_bio_iter() call.
> 
> In iomap_dio_bio_iter(), ensure that for a non-dsync iocb that the mapping
> is not dirty nor unmapped.
> 
> Signed-off-by: John Garry <john.g.garry at oracle.com>
> ---
>  fs/iomap/direct-io.c | 72 ++++++++++++++++++++++++++++++++++++++++++--
>  1 file changed, 70 insertions(+), 2 deletions(-)
> 
> diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
> index f771001574d0..37c3c926dfd8 100644
> --- a/fs/iomap/direct-io.c
> +++ b/fs/iomap/direct-io.c
> @@ -36,6 +36,8 @@ struct iomap_dio {
>  	size_t			done_before;
>  	bool			wait_for_completion;
>  
> +	unsigned int atomic_write_unit;
> +
>  	union {
>  		/* used during submission and for synchronous completion: */
>  		struct {
> @@ -229,9 +231,21 @@ static inline blk_opf_t iomap_dio_bio_opflags(struct iomap_dio *dio,
>  	return opflags;
>  }
>  
> +
> +/*
> + * Note: For atomic writes, each bio which we create when we iter should have
> + *	 bi_sector aligned to atomic_write_unit and also its bi_size should be
> + *	 a multiple of atomic_write_unit.
> + *	 The call to bio_iov_iter_get_pages() -> __bio_iov_iter_get_pages()
> + *	 should trim the length to a multiple of atomic_write_unit for us.
> + *	 This allows us to split each bio later in the block layer to fit
> + *	 request_queue limit.
> + */
>  static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
>  		struct iomap_dio *dio)
>  {
> +	bool atomic_write = (dio->iocb->ki_flags & IOCB_ATOMIC) &&
> +			    (dio->flags & IOMAP_DIO_WRITE);
>  	const struct iomap *iomap = &iter->iomap;
>  	struct inode *inode = iter->inode;
>  	unsigned int fs_block_size = i_blocksize(inode), pad;
> @@ -249,6 +263,14 @@ static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
>  	    !bdev_iter_is_aligned(iomap->bdev, dio->submit.iter))
>  		return -EINVAL;
>  
> +
> +	if (atomic_write && !iocb_is_dsync(dio->iocb)) {
> +		if (iomap->flags & IOMAP_F_DIRTY)
> +			return -EIO;
> +		if (iomap->type != IOMAP_MAPPED)
> +			return -EIO;
> +	}

IDGI. If the iomap had space allocated for this dio iteration,
then IOMAP_F_DIRTY will be set and it is likely (guaranteed for XFS)
that the iomap type will be IOMAP_UNWRITTEN. Indeed, if we are doing
a write into preallocated space (i.e. from fallocate()) then this
will cause -EIO on all RWF_ATOMIC IO to that file unless RWF_DSYNC
is also used.

"For a power fail, for each individual application block, all or
none of the data to be written."

Ok, does this means RWF_ATOMIC still needs fdatasync() to guarantee
that the data makes it to stable storage? And the result is
undefined until fdatasync() is run, but the device will guarantee
that either all or none of the data will be on stable storage
prior to the next device cache flush completing?

i.e. does REQ_ATOMIC imply REQ_FUA, or does it require a separate
device cache flush to commit the atomic IO to stable storage?

What about ordering - do the devices guarantee strict ordering of
REQ_ATOMIC writes? i.e. if atomic write N is seen on disk, then all
the previous atomic writes up to N will also be seen on disk? If
not, how does the application and filesystem guarantee persistence
of completed atomic writes?

i.e. If we still need a post-IO device cache flush to guarantee
persistence and/or ordering of RWF_ATOMIC IOs, then the above code
makes no sense - we'll still need fdatasync() to provide persistence
checkpoints and that means we ensure metadata is also up to date
at those checkpoints.

I need someone to put down in writing exactly what the data
integrity, ordering and persistence semantics of REQ_ATOMIC are
before I can really comment any further. From my perspective as a
filesystem developer, this is the single most important set of
behaviours that need to be documented, as this determines how
everything else interacts with atomic writes....

>  	if (iomap->type == IOMAP_UNWRITTEN) {
>  		dio->flags |= IOMAP_DIO_UNWRITTEN;
>  		need_zeroout = true;
> @@ -318,6 +340,10 @@ static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
>  					  GFP_KERNEL);
>  		bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
>  		bio->bi_ioprio = dio->iocb->ki_ioprio;
> +		if (atomic_write) {
> +			bio->bi_opf |= REQ_ATOMIC;
> +			bio->atomic_write_unit = dio->atomic_write_unit;
> +		}
>  		bio->bi_private = dio;
>  		bio->bi_end_io = iomap_dio_bio_end_io;
>  
> @@ -492,6 +518,8 @@ __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
>  		is_sync_kiocb(iocb) || (dio_flags & IOMAP_DIO_FORCE_WAIT);
>  	struct blk_plug plug;
>  	struct iomap_dio *dio;
> +	bool is_read = iov_iter_rw(iter) == READ;
> +	bool atomic_write = (iocb->ki_flags & IOCB_ATOMIC) && !is_read;
>  
>  	if (!iomi.len)
>  		return NULL;
> @@ -500,6 +528,20 @@ __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
>  	if (!dio)
>  		return ERR_PTR(-ENOMEM);
>  
> +	if (atomic_write) {
> +		/*
> +		 * Note: This lookup is not proper for a multi-device scenario,
> +		 *	 however for current iomap users, the bdev per iter
> +		 *	 will be fixed, so "works" for now.
> +		 */
> +		struct super_block *i_sb = inode->i_sb;
> +		struct block_device *bdev = i_sb->s_bdev;
> +
> +		dio->atomic_write_unit =
> +			bdev_find_max_atomic_write_alignment(bdev,
> +					iomi.pos, iomi.len);
> +	}

This will break atomic IO to XFS realtime devices. The device we are
doing IO to is iomap->bdev, we should never be using sb->s_bdev in
the iomap code.  Of course, at this point in __iomap_dio_rw() we
don't have an iomap so this "alignment constraint" can't be done
correctly at this point in the IO path.

However, even ignoring the bdev source, I think this is completely
wrong. Passing a *file* offset to the underlying block device so the
block device can return a device alignment constraint for IO is not
valid. We don't know how that file offset/length is going to be
mapped to the underlying block device until we ask the filesystem
for an iomap covering the file range, so we can't possibly know what
the device IO alignment of the user request will be until we have an
iomap for it.

At which point, the "which block device should we ask for alignment
constraints" question is moot, because we now have an iomap and can
use iomap->bdev....

> @@ -592,6 +634,32 @@ __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
>  
>  	blk_start_plug(&plug);
>  	while ((ret = iomap_iter(&iomi, ops)) > 0) {
> +		if (atomic_write) {
> +			const struct iomap *_iomap = &iomi.iomap;
> +			loff_t iomi_length = iomap_length(&iomi);
> +
> +			/*
> +			 * Ensure length and start address is a multiple of
> +			 * atomic_write_unit - this is critical. If the length
> +			 * is not a multiple of atomic_write_unit, then we
> +			 * cannot create a set of bio's in iomap_dio_bio_iter()
> +			 * who are each a length which is a multiple of
> +			 * atomic_write_unit.
> +			 *
> +			 * Note: It may be more appropiate to have this check
> +			 *	 in iomap_dio_bio_iter()
> +			 */
> +			if ((iomap_sector(_iomap, iomi.pos) << SECTOR_SHIFT) %
> +			    dio->atomic_write_unit) {
> +				ret = -EIO;
> +				break;
> +			}
> +
> +			if (iomi_length % dio->atomic_write_unit) {
> +				ret = -EIO;
> +				break;
> +			}

This looks wrong - the length of the mapped extent could be shorter
than the max atomic write size returned by
bdev_find_max_atomic_write_alignment() but the iomap could still be aligned
to the minimum atomic write unit supported. At this point, we reject
the IO with -EIO, even though it could have been done as an atomic
write, just a shorter one than the user requested.

That said, I don't think we can call a user IO that is being
sliced and diced into multiple individual IOs "atomic". "Atomic"
implies all-or-none behaviour - slicing up a large DIO into smaller
individual bios means the bios can be submitted and completed out of
order. If we then we get a power failure, the application's "atomic"
IO can appear on disk as only being partially complete - it violates
the "all or none" semantics of "atomic IO".

Hence I think that we should be rejecting RWF_ATOMIC IOs that are
larger than the maximum atomic write unit or cannot be dispatched in
a single IO e.g. filesystem has allocated multiple minimum aligned
extents and so a max len atomic write IO over that range must be
broken up into multiple smaller IOs.

We should be doing max atomic write size rejection high up in the IO
path (e.g. filesystem ->write_iter() method) before we get anywhere
near the DIO path, and we should be rejecting atomic write IOs in
the DIO path during the ->iomap_begin() mapping callback if we can't
map the entire atomic IO to a single aligned filesystem extent.

i.e. the alignment checks and constraints need to be applied by the
filesystem mapping code, not the layer that packs the pages into the
bio as directed by the filesystem mapping....

Cheers,

Dave.
-- 
Dave Chinner
david at fromorbit.com



More information about the Linux-nvme mailing list