Re: [PATCH v3 5/7] block: fail atomic writes instead of falling back to buffered I/O

[John Garry <john.garry@linux.dev>]

On 9/9/26 23:05, Tal Zussman wrote:
> An IOCB_ATOMIC direct write to a block device can silently lose its
> torn-write guarantee in two ways:
> 
>    1. blkdev_direct_write() turns an -EBUSY from page cache invalidation
>       into a 0 return, so the whole write is retried through
>       blkdev_buffered_write(), with no atomicity guarantee.
> 
>    2. On a partial page pin, __blkdev_direct_IO_simple() and
>       __blkdev_direct_IO_async() submit what was pinned with REQ_ATOMIC
>       set and leave the rest to the buffered fallback.

This really should be 2x separate changes - 1x for fops.c and 1x for bio.c

> 
> The second case can be triggered deterministically. A 16K
> pwritev2(RWF_ATOMIC) whose last page is PROT_NONE, on a scsi_debug
> device with atomic_wr=1, completes short with only three of the four
> pages written, violating RWF_ATOMIC semantics.
> 
> Fail the I/O instead. Make bio_iov_iter_get_pages() release the pins
> and return -EINVAL when a REQ_ATOMIC bio doesn't cover the whole
> iterator, since an atomic write is submitted as a single bio and a
> short one would be torn. That covers iomap as well, where a partially
> unmapped buffer could trip the WARN_ON_ONCE() in
> iomap_dio_bio_iter_one(). The async block device path currently sets
> REQ_ATOMIC after pinning, so set it before.
> 
> Skip the buffered fallback in blkdev_write_iter() for IOCB_ATOMIC, as
> it already does for IOCB_NOWAIT, so the -EBUSY case returns -EAGAIN and
> the caller retries, matching __iomap_dio_rw().
> 
> ext4 has the same fallback and only warns in it. For block devices both
> ways in can be detected before any I/O is submitted, so fail early instead.
> 
> Fixes: caf336f81b3a ("block: Add fops atomic write support")
> Reported-by: Sashiko <sashiko-bot@kernel.org>
> Link: https://sashiko.dev/#/patchset/20260802-blkdev-fixes-v1-0-a82fc549fd74%40columbia.edu?part=2
> Assisted-by: Claude:claude-fable-5
> Signed-off-by: Tal Zussman <tz2294@columbia.edu>
> ---
>   block/bio.c  | 29 ++++++++++++++++++++++-------
>   block/fops.c | 10 +++++-----
>   2 files changed, 27 insertions(+), 12 deletions(-)
> 
> diff --git a/block/bio.c b/block/bio.c
> index 898b2f5ef8c8..63e266d861f1 100644
> --- a/block/bio.c
> +++ b/block/bio.c
> @@ -1284,6 +1284,7 @@ int bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter,
>   			   unsigned mem_align_mask, unsigned len_align_mask)
>   {
>   	iov_iter_extraction_t flags = 0;
> +	int ret;
>   
>   	if (WARN_ON_ONCE(bio_flagged(bio, BIO_CLONED)))
>   		return -EIO;
> @@ -1303,34 +1304,48 @@ int bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter,
>   		flags |= ITER_ALLOW_P2PDMA;
>   
>   	do {
> -		ssize_t ret;
> +		ssize_t len;

iov_iter_extract_bvecs() local variable is called "size", so maybe use 
the same here

>   
> -		ret = iov_iter_extract_bvecs(iter, bio->bi_io_vec,
> +		len = iov_iter_extract_bvecs(iter, bio->bi_io_vec,
>   				BIO_MAX_SIZE - bio->bi_iter.bi_size,
>   				&bio->bi_vcnt, bio->bi_max_vecs,
>   				mem_align_mask, flags);
> -		if (ret <= 0) {
> +		if (len <= 0) {
>   			/*
>   			 * A misaligned vector fails the whole I/O.  Release any
>   			 * pages pinned by earlier iterations before returning
>   			 * since this bio won't be submitted to release them.
>   			 */
> -			if (ret == -EINVAL) {
> +			if (len == -EINVAL) {
>   				bio_release_pages(bio, false);
>   				bio_clear_flag(bio, BIO_PAGE_PINNED);
>   				bio->bi_vcnt = 0;
>   			}
>   			if (!bio->bi_vcnt)
> -				return ret;
> +				return len;
>   			break;
>   		}
> -		bio->bi_iter.bi_size += ret;
> +		bio->bi_iter.bi_size += len;
>   	} while (iov_iter_count(iter) && !bio_full(bio, 0));
>   
>   	if (is_pci_p2pdma_page(bio->bi_io_vec->bv_page))
>   		bio->bi_opf |= REQ_NOMERGE;
> -	return bio_iov_iter_align_down(bio, iter,
> +	ret = bio_iov_iter_align_down(bio, iter,
>   			&bio->bi_io_vec[bio->bi_vcnt - 1], len_align_mask);
> +	if (ret)
> +		return ret;
 > +> +	/*
> +	 * An atomic write is submitted as a single bio, so it has to cover
> +	 * the whole iterator or it would be torn.
> +	 */
> +	if ((bio->bi_opf & REQ_ATOMIC) && iov_iter_count(iter)) {
> +		bio_release_pages(bio, false);
> +		bio_clear_flag(bio, BIO_PAGE_PINNED);
> +		bio->bi_vcnt = 0;
> +		return -EINVAL;
> +	}

This all looks ok, but I'll check again ...

> +	return 0;
>   }

iomap_dio_bio_iter_one() can be updated at some stage to remove its own 
check for improper length returned from bio_iov_iter_get_pages() for 
IOCB_ATOMIC

>   
>   static struct folio *folio_alloc_greedy(gfp_t gfp, size_t *size,
> diff --git a/block/fops.c b/block/fops.c
> index a3a709697b40..0b614d76d128 100644
> --- a/block/fops.c
> +++ b/block/fops.c
> @@ -341,6 +341,8 @@ static ssize_t __blkdev_direct_IO_async(struct kiocb *iocb,
>   	bio->bi_write_stream = iocb->ki_write_stream;
>   	bio->bi_end_io = blkdev_bio_end_io_async;
>   	bio->bi_ioprio = iocb->ki_ioprio;
> +	if (iocb->ki_flags & IOCB_ATOMIC)
> +		bio->bi_opf |= REQ_ATOMIC;
>   
>   	/*
>   	 * Users don't rely on the iterator being in any particular
> @@ -371,9 +373,6 @@ static ssize_t __blkdev_direct_IO_async(struct kiocb *iocb,
>   			goto out_bio_put;
>   	}
>   
> -	if (iocb->ki_flags & IOCB_ATOMIC)
> -		bio->bi_opf |= REQ_ATOMIC;
> -
>   	if (iocb->ki_flags & IOCB_NOWAIT)
>   		bio->bi_opf |= REQ_NOWAIT;

I think that you relocate this as well to have similar functionality 
co-located

>   
> @@ -766,10 +765,11 @@ static ssize_t blkdev_write_iter(struct kiocb *iocb, struct iov_iter *from)
>   	if (iocb->ki_flags & IOCB_DIRECT) {
>   		ret = blkdev_direct_write(iocb, from);
>   		if (ret >= 0 && iov_iter_count(from)) {
> -			if (iocb->ki_flags & IOCB_NOWAIT) {
> +			if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_ATOMIC)) {

An alternative could be to have iomap_file_buffered_write() reject 
IOCB_ATOMIC.

>   				/*
>   				 * The buffered fallback blocks on i_rwsem and
> -				 * on writeback of the data it copied: return
> +				 * on writeback of the data it copied, and
> +				 * can't provide torn-write protection: return
>   				 * the short direct write instead and let the
>   				 * caller retry.
>   				 */