bounce buffer direct I/O when stable pages are required v3

bounce buffer direct I/O when stable pages are required v3

[Christoph Hellwig]

Hi all,

[note to maintainers:  we're ready to merge I think, and Christian
already said he'd do on Friday.  If acceptable to everyone I'd like
to merge it through the block tree, or topic branch in it due to
pending work on top of this]

this series tries to address the problem that under I/O pages can be
modified during direct I/O, even when the device or file system require
stable pages during I/O to calculate checksums, parity or data
operations.  It does so by adding block layer helpers to bounce buffer
an iov_iter into a bio, then wires that up in iomap and ultimately
XFS.

The reason that the file system even needs to know about it, is because
reads need a user context to copy the data back, and the infrastructure
to defer ioends to a workqueue currently sits in XFS.  I'm going to look
into moving that into ioend and enabling it for other file systems.
Additionally btrfs already has it's own infrastructure for this, and
actually an urgent need to bounce buffer, so this should be useful there
and could be wire up easily.  In fact the idea comes from patches by
Qu that did this in btrfs.

This patch fixes all but one xfstests failures on T10 PI capable devices
(generic/095 seems to have issues with a mix of mmap and splice still,
I'm looking into that separate), and make qemu VMs running Windows,
or Linux with swap enabled fine on an XFS file on a device using PI.

Performance numbers on my (not exactly state of the art) NVMe PI test
setup:

  Sequential reads using io_uring, QD=16.
  Bandwidth and CPU usage (usr/sys):

  | size |        zero copy         |          bounce          |
  +------+--------------------------+--------------------------+
  |   4k | 1316MiB/s (12.65/55.40%) | 1081MiB/s (11.76/49.78%) |
  |  64K | 3370MiB/s ( 5.46/18.20%) | 3365MiB/s ( 4.47/15.68%) |
  |   1M | 3401MiB/s ( 0.76/23.05%) | 3400MiB/s ( 0.80/09.06%) |
  +------+--------------------------+--------------------------+

  Sequential writes using io_uring, QD=16.
  Bandwidth and CPU usage (usr/sys):

  | size |        zero copy         |          bounce          |
  +------+--------------------------+--------------------------+
  |   4k |  882MiB/s (11.83/33.88%) |  750MiB/s (10.53/34.08%) |
  |  64K | 2009MiB/s ( 7.33/15.80%) | 2007MiB/s ( 7.47/24.71%) |
  |   1M | 1992MiB/s ( 7.26/ 9.13%) | 1992MiB/s ( 9.21/19.11%) |
  +------+--------------------------+--------------------------+

Note that the 64k read numbers look really odd to me for the baseline
zero copy case, but are reproducible over many repeated runs.

The bounce read numbers should further improve when moving the PI
validation to the file system and removing the double context switch,
which I have patches for that will sent out soon.

Changes since v2:
 - add a BIO_MAX_SIZE constant and use it
 - remove a pointless repeated page_folio call
 - fix a comment typo
 - add a new comment about copying to a pinned iter

Changes since v1:
 - spelling fixes
 - add more details to some commit messages
 - add a new code comment about freeing the bio early in the I/O
   completion handler

Diffstat:
 block/bio.c               |  332 ++++++++++++++++++++++++++++------------------
 block/blk-lib.c           |    9 -
 block/blk-merge.c         |    8 -
 block/blk.h               |   11 -
 fs/iomap/direct-io.c      |  191 ++++++++++++++------------
 fs/iomap/ioend.c          |    8 +
 fs/xfs/xfs_aops.c         |    8 -
 fs/xfs/xfs_file.c         |   41 +++++
 include/linux/bio.h       |   26 +++
 include/linux/blk_types.h |    3 
 include/linux/iomap.h     |    9 +
 include/linux/uio.h       |    3 
 lib/iov_iter.c            |   98 +++++++++++++
 13 files changed, 507 insertions(+), 240 deletions(-)

[PATCH 01/15] block: add a BIO_MAX_SIZE constant and use it

[Christoph Hellwig]

Currently the only constant for the maximum bio size is BIO_MAX_SECTORS,
which is in units of 512-byte sectors, but a lot of user need a byte
limit.

Add a BIO_MAX_SIZE constant, redefine BIO_MAX_SECTORS in terms of it, and
switch all bio-related uses of UINT_MAX for the maximum size to use the
symbolic names instead.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 block/bio.c               | 10 +++++-----
 block/blk-lib.c           |  9 ++++-----
 block/blk-merge.c         |  8 ++++----
 include/linux/blk_types.h |  3 ++-
 4 files changed, 15 insertions(+), 15 deletions(-)

diff --git a/block/bio.c b/block/bio.c
index 2359c0723b88..ac7703e149c6 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -924,7 +924,7 @@ static inline bool bio_full(struct bio *bio, unsigned len)
 {
 	if (bio->bi_vcnt >= bio->bi_max_vecs)
 		return true;
-	if (bio->bi_iter.bi_size > UINT_MAX - len)
+	if (bio->bi_iter.bi_size > BIO_MAX_SIZE - len)
 		return true;
 	return false;
 }
@@ -1030,7 +1030,7 @@ int bio_add_page(struct bio *bio, struct page *page,
 {
 	if (WARN_ON_ONCE(bio_flagged(bio, BIO_CLONED)))
 		return 0;
-	if (bio->bi_iter.bi_size > UINT_MAX - len)
+	if (bio->bi_iter.bi_size > BIO_MAX_SIZE - len)
 		return 0;
 
 	if (bio->bi_vcnt > 0) {
@@ -1057,7 +1057,7 @@ void bio_add_folio_nofail(struct bio *bio, struct folio *folio, size_t len,
 {
 	unsigned long nr = off / PAGE_SIZE;
 
-	WARN_ON_ONCE(len > UINT_MAX);
+	WARN_ON_ONCE(len > BIO_MAX_SIZE);
 	__bio_add_page(bio, folio_page(folio, nr), len, off % PAGE_SIZE);
 }
 EXPORT_SYMBOL_GPL(bio_add_folio_nofail);
@@ -1081,7 +1081,7 @@ bool bio_add_folio(struct bio *bio, struct folio *folio, size_t len,
 {
 	unsigned long nr = off / PAGE_SIZE;
 
-	if (len > UINT_MAX)
+	if (len > BIO_MAX_SIZE)
 		return false;
 	return bio_add_page(bio, folio_page(folio, nr), len, off % PAGE_SIZE) > 0;
 }
@@ -1238,7 +1238,7 @@ static int __bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
 		extraction_flags |= ITER_ALLOW_P2PDMA;
 
 	size = iov_iter_extract_pages(iter, &pages,
-				      UINT_MAX - bio->bi_iter.bi_size,
+				      BIO_MAX_SIZE - bio->bi_iter.bi_size,
 				      nr_pages, extraction_flags, &offset);
 	if (unlikely(size <= 0))
 		return size ? size : -EFAULT;
diff --git a/block/blk-lib.c b/block/blk-lib.c
index 9e2cc58f881f..0be3acdc3eb5 100644
--- a/block/blk-lib.c
+++ b/block/blk-lib.c
@@ -32,7 +32,7 @@ static sector_t bio_discard_limit(struct block_device *bdev, sector_t sector)
 	 * Align the bio size to the discard granularity to make splitting the bio
 	 * at discard granularity boundaries easier in the driver if needed.
 	 */
-	return round_down(UINT_MAX, discard_granularity) >> SECTOR_SHIFT;
+	return round_down(BIO_MAX_SIZE, discard_granularity) >> SECTOR_SHIFT;
 }
 
 struct bio *blk_alloc_discard_bio(struct block_device *bdev,
@@ -107,8 +107,7 @@ static sector_t bio_write_zeroes_limit(struct block_device *bdev)
 {
 	sector_t bs_mask = (bdev_logical_block_size(bdev) >> 9) - 1;
 
-	return min(bdev_write_zeroes_sectors(bdev),
-		(UINT_MAX >> SECTOR_SHIFT) & ~bs_mask);
+	return min(bdev_write_zeroes_sectors(bdev), BIO_MAX_SECTORS & ~bs_mask);
 }
 
 /*
@@ -337,8 +336,8 @@ int blkdev_issue_secure_erase(struct block_device *bdev, sector_t sector,
 	int ret = 0;
 
 	/* make sure that "len << SECTOR_SHIFT" doesn't overflow */
-	if (max_sectors > UINT_MAX >> SECTOR_SHIFT)
-		max_sectors = UINT_MAX >> SECTOR_SHIFT;
+	if (max_sectors > BIO_MAX_SECTORS)
+		max_sectors = BIO_MAX_SECTORS;
 	max_sectors &= ~bs_mask;
 
 	if (max_sectors == 0)
diff --git a/block/blk-merge.c b/block/blk-merge.c
index b82c6d304658..0eb0aef97197 100644
--- a/block/blk-merge.c
+++ b/block/blk-merge.c
@@ -95,13 +95,13 @@ static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
 }
 
 /*
- * The max size one bio can handle is UINT_MAX becasue bvec_iter.bi_size
- * is defined as 'unsigned int', meantime it has to be aligned to with the
+ * The maximum size that a bio can fit has to be aligned down to the
  * logical block size, which is the minimum accepted unit by hardware.
  */
 static unsigned int bio_allowed_max_sectors(const struct queue_limits *lim)
 {
-	return round_down(UINT_MAX, lim->logical_block_size) >> SECTOR_SHIFT;
+	return round_down(BIO_MAX_SIZE, lim->logical_block_size) >>
+			SECTOR_SHIFT;
 }
 
 /*
@@ -502,7 +502,7 @@ unsigned int blk_recalc_rq_segments(struct request *rq)
 
 	rq_for_each_bvec(bv, rq, iter)
 		bvec_split_segs(&rq->q->limits, &bv, &nr_phys_segs, &bytes,
-				UINT_MAX, UINT_MAX);
+				UINT_MAX, BIO_MAX_SIZE);
 	return nr_phys_segs;
 }
 
diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
index 19a888a2f104..d59553324a84 100644
--- a/include/linux/blk_types.h
+++ b/include/linux/blk_types.h
@@ -281,7 +281,8 @@ struct bio {
 };
 
 #define BIO_RESET_BYTES		offsetof(struct bio, bi_max_vecs)
-#define BIO_MAX_SECTORS		(UINT_MAX >> SECTOR_SHIFT)
+#define BIO_MAX_SIZE		UINT_MAX /* max value of bi_iter.bi_size */
+#define BIO_MAX_SECTORS		(BIO_MAX_SIZE >> SECTOR_SHIFT)
 
 static inline struct bio_vec *bio_inline_vecs(struct bio *bio)
 {
-- 
2.47.3

Re: [PATCH 01/15] block: add a BIO_MAX_SIZE constant and use it

[Damien Le Moal]

On 1/26/26 2:53 PM, Christoph Hellwig wrote:
> Currently the only constant for the maximum bio size is BIO_MAX_SECTORS,
> which is in units of 512-byte sectors, but a lot of user need a byte
> limit.
> 
> Add a BIO_MAX_SIZE constant, redefine BIO_MAX_SECTORS in terms of it, and
> switch all bio-related uses of UINT_MAX for the maximum size to use the
> symbolic names instead.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>

Nice cleanup !

Reviewed-by: Damien Le Moal <dlemoal@kernel.org>

-- 
Damien Le Moal
Western Digital Research

Re: [PATCH 01/15] block: add a BIO_MAX_SIZE constant and use it

[Johannes Thumshirn]

Looks good,

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>

Re: [PATCH 01/15] block: add a BIO_MAX_SIZE constant and use it

[Anuj gupta]

Reviewed-by: Anuj Gupta <anuj20.g@samsung.com>

Re: [PATCH 01/15] block: add a BIO_MAX_SIZE constant and use it

[Darrick J. Wong]

On Mon, Jan 26, 2026 at 06:53:32AM +0100, Christoph Hellwig wrote:
> Currently the only constant for the maximum bio size is BIO_MAX_SECTORS,
> which is in units of 512-byte sectors, but a lot of user need a byte
> limit.
> 
> Add a BIO_MAX_SIZE constant, redefine BIO_MAX_SECTORS in terms of it, and
> switch all bio-related uses of UINT_MAX for the maximum size to use the
> symbolic names instead.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>

Yay, thanks!
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>

--D

> ---
>  block/bio.c               | 10 +++++-----
>  block/blk-lib.c           |  9 ++++-----
>  block/blk-merge.c         |  8 ++++----
>  include/linux/blk_types.h |  3 ++-
>  4 files changed, 15 insertions(+), 15 deletions(-)
> 
> diff --git a/block/bio.c b/block/bio.c
> index 2359c0723b88..ac7703e149c6 100644
> --- a/block/bio.c
> +++ b/block/bio.c
> @@ -924,7 +924,7 @@ static inline bool bio_full(struct bio *bio, unsigned len)
>  {
>  	if (bio->bi_vcnt >= bio->bi_max_vecs)
>  		return true;
> -	if (bio->bi_iter.bi_size > UINT_MAX - len)
> +	if (bio->bi_iter.bi_size > BIO_MAX_SIZE - len)
>  		return true;
>  	return false;
>  }
> @@ -1030,7 +1030,7 @@ int bio_add_page(struct bio *bio, struct page *page,
>  {
>  	if (WARN_ON_ONCE(bio_flagged(bio, BIO_CLONED)))
>  		return 0;
> -	if (bio->bi_iter.bi_size > UINT_MAX - len)
> +	if (bio->bi_iter.bi_size > BIO_MAX_SIZE - len)
>  		return 0;
>  
>  	if (bio->bi_vcnt > 0) {
> @@ -1057,7 +1057,7 @@ void bio_add_folio_nofail(struct bio *bio, struct folio *folio, size_t len,
>  {
>  	unsigned long nr = off / PAGE_SIZE;
>  
> -	WARN_ON_ONCE(len > UINT_MAX);
> +	WARN_ON_ONCE(len > BIO_MAX_SIZE);
>  	__bio_add_page(bio, folio_page(folio, nr), len, off % PAGE_SIZE);
>  }
>  EXPORT_SYMBOL_GPL(bio_add_folio_nofail);
> @@ -1081,7 +1081,7 @@ bool bio_add_folio(struct bio *bio, struct folio *folio, size_t len,
>  {
>  	unsigned long nr = off / PAGE_SIZE;
>  
> -	if (len > UINT_MAX)
> +	if (len > BIO_MAX_SIZE)
>  		return false;
>  	return bio_add_page(bio, folio_page(folio, nr), len, off % PAGE_SIZE) > 0;
>  }
> @@ -1238,7 +1238,7 @@ static int __bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
>  		extraction_flags |= ITER_ALLOW_P2PDMA;
>  
>  	size = iov_iter_extract_pages(iter, &pages,
> -				      UINT_MAX - bio->bi_iter.bi_size,
> +				      BIO_MAX_SIZE - bio->bi_iter.bi_size,
>  				      nr_pages, extraction_flags, &offset);
>  	if (unlikely(size <= 0))
>  		return size ? size : -EFAULT;
> diff --git a/block/blk-lib.c b/block/blk-lib.c
> index 9e2cc58f881f..0be3acdc3eb5 100644
> --- a/block/blk-lib.c
> +++ b/block/blk-lib.c
> @@ -32,7 +32,7 @@ static sector_t bio_discard_limit(struct block_device *bdev, sector_t sector)
>  	 * Align the bio size to the discard granularity to make splitting the bio
>  	 * at discard granularity boundaries easier in the driver if needed.
>  	 */
> -	return round_down(UINT_MAX, discard_granularity) >> SECTOR_SHIFT;
> +	return round_down(BIO_MAX_SIZE, discard_granularity) >> SECTOR_SHIFT;
>  }
>  
>  struct bio *blk_alloc_discard_bio(struct block_device *bdev,
> @@ -107,8 +107,7 @@ static sector_t bio_write_zeroes_limit(struct block_device *bdev)
>  {
>  	sector_t bs_mask = (bdev_logical_block_size(bdev) >> 9) - 1;
>  
> -	return min(bdev_write_zeroes_sectors(bdev),
> -		(UINT_MAX >> SECTOR_SHIFT) & ~bs_mask);
> +	return min(bdev_write_zeroes_sectors(bdev), BIO_MAX_SECTORS & ~bs_mask);
>  }
>  
>  /*
> @@ -337,8 +336,8 @@ int blkdev_issue_secure_erase(struct block_device *bdev, sector_t sector,
>  	int ret = 0;
>  
>  	/* make sure that "len << SECTOR_SHIFT" doesn't overflow */
> -	if (max_sectors > UINT_MAX >> SECTOR_SHIFT)
> -		max_sectors = UINT_MAX >> SECTOR_SHIFT;
> +	if (max_sectors > BIO_MAX_SECTORS)
> +		max_sectors = BIO_MAX_SECTORS;
>  	max_sectors &= ~bs_mask;
>  
>  	if (max_sectors == 0)
> diff --git a/block/blk-merge.c b/block/blk-merge.c
> index b82c6d304658..0eb0aef97197 100644
> --- a/block/blk-merge.c
> +++ b/block/blk-merge.c
> @@ -95,13 +95,13 @@ static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
>  }
>  
>  /*
> - * The max size one bio can handle is UINT_MAX becasue bvec_iter.bi_size
> - * is defined as 'unsigned int', meantime it has to be aligned to with the
> + * The maximum size that a bio can fit has to be aligned down to the
>   * logical block size, which is the minimum accepted unit by hardware.
>   */
>  static unsigned int bio_allowed_max_sectors(const struct queue_limits *lim)
>  {
> -	return round_down(UINT_MAX, lim->logical_block_size) >> SECTOR_SHIFT;
> +	return round_down(BIO_MAX_SIZE, lim->logical_block_size) >>
> +			SECTOR_SHIFT;
>  }
>  
>  /*
> @@ -502,7 +502,7 @@ unsigned int blk_recalc_rq_segments(struct request *rq)
>  
>  	rq_for_each_bvec(bv, rq, iter)
>  		bvec_split_segs(&rq->q->limits, &bv, &nr_phys_segs, &bytes,
> -				UINT_MAX, UINT_MAX);
> +				UINT_MAX, BIO_MAX_SIZE);
>  	return nr_phys_segs;
>  }
>  
> diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
> index 19a888a2f104..d59553324a84 100644
> --- a/include/linux/blk_types.h
> +++ b/include/linux/blk_types.h
> @@ -281,7 +281,8 @@ struct bio {
>  };
>  
>  #define BIO_RESET_BYTES		offsetof(struct bio, bi_max_vecs)
> -#define BIO_MAX_SECTORS		(UINT_MAX >> SECTOR_SHIFT)
> +#define BIO_MAX_SIZE		UINT_MAX /* max value of bi_iter.bi_size */
> +#define BIO_MAX_SECTORS		(BIO_MAX_SIZE >> SECTOR_SHIFT)
>  
>  static inline struct bio_vec *bio_inline_vecs(struct bio *bio)
>  {
> -- 
> 2.47.3
> 
> 

Re: [PATCH 01/15] block: add a BIO_MAX_SIZE constant and use it

[Martin K. Petersen]

Christoph,

> Currently the only constant for the maximum bio size is
> BIO_MAX_SECTORS, which is in units of 512-byte sectors, but a lot of
> user need a byte limit.
>
> Add a BIO_MAX_SIZE constant, redefine BIO_MAX_SECTORS in terms of it,
> and switch all bio-related uses of UINT_MAX for the maximum size to
> use the symbolic names instead.

Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>

-- 
Martin K. Petersen

[PATCH 02/15] block: refactor get_contig_folio_len

[Christoph Hellwig]

Move all of the logic to find the contigous length inside a folio into
get_contig_folio_len instead of keeping some of it in the caller.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
---
 block/bio.c | 62 +++++++++++++++++++++++------------------------------
 1 file changed, 27 insertions(+), 35 deletions(-)

diff --git a/block/bio.c b/block/bio.c
index ac7703e149c6..d633e80d821f 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -1172,33 +1172,35 @@ void bio_iov_bvec_set(struct bio *bio, const struct iov_iter *iter)
 	bio_set_flag(bio, BIO_CLONED);
 }
 
-static unsigned int get_contig_folio_len(unsigned int *num_pages,
-					 struct page **pages, unsigned int i,
-					 struct folio *folio, size_t left,
+static unsigned int get_contig_folio_len(struct page **pages,
+					 unsigned int *num_pages, size_t left,
 					 size_t offset)
 {
-	size_t bytes = left;
-	size_t contig_sz = min_t(size_t, PAGE_SIZE - offset, bytes);
-	unsigned int j;
+	struct folio *folio = page_folio(pages[0]);
+	size_t contig_sz = min_t(size_t, PAGE_SIZE - offset, left);
+	unsigned int max_pages, i;
+	size_t folio_offset, len;
+
+	folio_offset = PAGE_SIZE * folio_page_idx(folio, pages[0]) + offset;
+	len = min(folio_size(folio) - folio_offset, left);
 
 	/*
-	 * We might COW a single page in the middle of
-	 * a large folio, so we have to check that all
-	 * pages belong to the same folio.
+	 * We might COW a single page in the middle of a large folio, so we have
+	 * to check that all pages belong to the same folio.
 	 */
-	bytes -= contig_sz;
-	for (j = i + 1; j < i + *num_pages; j++) {
-		size_t next = min_t(size_t, PAGE_SIZE, bytes);
+	left -= contig_sz;
+	max_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
+	for (i = 1; i < max_pages; i++) {
+		size_t next = min_t(size_t, PAGE_SIZE, left);
 
-		if (page_folio(pages[j]) != folio ||
-		    pages[j] != pages[j - 1] + 1) {
+		if (page_folio(pages[i]) != folio ||
+		    pages[i] != pages[i - 1] + 1)
 			break;
-		}
 		contig_sz += next;
-		bytes -= next;
+		left -= next;
 	}
-	*num_pages = j - i;
 
+	*num_pages = i;
 	return contig_sz;
 }
 
@@ -1222,8 +1224,8 @@ static int __bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
 	struct bio_vec *bv = bio->bi_io_vec + bio->bi_vcnt;
 	struct page **pages = (struct page **)bv;
 	ssize_t size;
-	unsigned int num_pages, i = 0;
-	size_t offset, folio_offset, left, len;
+	unsigned int i = 0;
+	size_t offset, left, len;
 	int ret = 0;
 
 	/*
@@ -1244,23 +1246,12 @@ static int __bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
 		return size ? size : -EFAULT;
 
 	nr_pages = DIV_ROUND_UP(offset + size, PAGE_SIZE);
-	for (left = size, i = 0; left > 0; left -= len, i += num_pages) {
-		struct page *page = pages[i];
-		struct folio *folio = page_folio(page);
+	for (left = size; left > 0; left -= len) {
 		unsigned int old_vcnt = bio->bi_vcnt;
+		unsigned int nr_to_add;
 
-		folio_offset = ((size_t)folio_page_idx(folio, page) <<
-			       PAGE_SHIFT) + offset;
-
-		len = min(folio_size(folio) - folio_offset, left);
-
-		num_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
-
-		if (num_pages > 1)
-			len = get_contig_folio_len(&num_pages, pages, i,
-						   folio, left, offset);
-
-		if (!bio_add_folio(bio, folio, len, folio_offset)) {
+		len = get_contig_folio_len(&pages[i], &nr_to_add, left, offset);
+		if (!bio_add_page(bio, pages[i], len, offset)) {
 			WARN_ON_ONCE(1);
 			ret = -EINVAL;
 			goto out;
@@ -1275,8 +1266,9 @@ static int __bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
 			 * single pin per page.
 			 */
 			if (offset && bio->bi_vcnt == old_vcnt)
-				unpin_user_folio(folio, 1);
+				unpin_user_folio(page_folio(pages[i]), 1);
 		}
+		i += nr_to_add;
 		offset = 0;
 	}
 
-- 
2.47.3

Re: [PATCH 02/15] block: refactor get_contig_folio_len

[Anuj gupta]

I sent this on the previous version, but FWIW
Reviewed-by: Anuj Gupta <anuj20.g@samsung.com>

Re: [PATCH 02/15] block: refactor get_contig_folio_len

[Martin K. Petersen]

Christoph,

> Move all of the logic to find the contigous length inside a folio into
> get_contig_folio_len instead of keeping some of it in the caller.

Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>

-- 
Martin K. Petersen

[PATCH 03/15] block: open code bio_add_page and fix handling of mismatching P2P ranges

[Christoph Hellwig]

bio_add_page fails to add data to the bio when mixing P2P with non-P2P
ranges, or ranges that map to different P2P providers.  In that case
it will trigger that WARN_ON and return an error up the chain instead of
simply starting a new bio as intended.  Fix this by open coding
bio_add_page and handling this case explicitly.  While doing so, stop
merging physical contiguous data that belongs to multiple folios.  While
this merge could lead to more efficient bio packing in some case,
dropping will allow to remove handling of this corner case in other
places and make the code more robust.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Anuj Gupta <anuj20.g@samsung.com>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
---
 block/bio.c | 37 +++++++++++++------------------------
 1 file changed, 13 insertions(+), 24 deletions(-)

diff --git a/block/bio.c b/block/bio.c
index d633e80d821f..4591f0ba90f5 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -1216,7 +1216,7 @@ static unsigned int get_contig_folio_len(struct page **pages,
  * For a multi-segment *iter, this function only adds pages from the next
  * non-empty segment of the iov iterator.
  */
-static int __bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
+static ssize_t __bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
 {
 	iov_iter_extraction_t extraction_flags = 0;
 	unsigned short nr_pages = bio->bi_max_vecs - bio->bi_vcnt;
@@ -1226,7 +1226,6 @@ static int __bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
 	ssize_t size;
 	unsigned int i = 0;
 	size_t offset, left, len;
-	int ret = 0;
 
 	/*
 	 * Move page array up in the allocated memory for the bio vecs as far as
@@ -1247,37 +1246,26 @@ static int __bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
 
 	nr_pages = DIV_ROUND_UP(offset + size, PAGE_SIZE);
 	for (left = size; left > 0; left -= len) {
-		unsigned int old_vcnt = bio->bi_vcnt;
 		unsigned int nr_to_add;
 
-		len = get_contig_folio_len(&pages[i], &nr_to_add, left, offset);
-		if (!bio_add_page(bio, pages[i], len, offset)) {
-			WARN_ON_ONCE(1);
-			ret = -EINVAL;
-			goto out;
-		}
+		if (bio->bi_vcnt > 0) {
+			struct bio_vec *prev = &bio->bi_io_vec[bio->bi_vcnt - 1];
 
-		if (bio_flagged(bio, BIO_PAGE_PINNED)) {
-			/*
-			 * We're adding another fragment of a page that already
-			 * was part of the last segment.  Undo our pin as the
-			 * page was pinned when an earlier fragment of it was
-			 * added to the bio and __bio_release_pages expects a
-			 * single pin per page.
-			 */
-			if (offset && bio->bi_vcnt == old_vcnt)
-				unpin_user_folio(page_folio(pages[i]), 1);
+			if (!zone_device_pages_have_same_pgmap(prev->bv_page,
+					pages[i]))
+				break;
 		}
+
+		len = get_contig_folio_len(&pages[i], &nr_to_add, left, offset);
+		__bio_add_page(bio, pages[i], len, offset);
 		i += nr_to_add;
 		offset = 0;
 	}
 
 	iov_iter_revert(iter, left);
-out:
 	while (i < nr_pages)
 		bio_release_page(bio, pages[i++]);
-
-	return ret;
+	return size - left;
 }
 
 /*
@@ -1337,7 +1325,7 @@ static int bio_iov_iter_align_down(struct bio *bio, struct iov_iter *iter,
 int bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter,
 			   unsigned len_align_mask)
 {
-	int ret = 0;
+	ssize_t ret;
 
 	if (WARN_ON_ONCE(bio_flagged(bio, BIO_CLONED)))
 		return -EIO;
@@ -1350,9 +1338,10 @@ int bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter,
 
 	if (iov_iter_extract_will_pin(iter))
 		bio_set_flag(bio, BIO_PAGE_PINNED);
+
 	do {
 		ret = __bio_iov_iter_get_pages(bio, iter);
-	} while (!ret && iov_iter_count(iter) && !bio_full(bio, 0));
+	} while (ret > 0 && iov_iter_count(iter) && !bio_full(bio, 0));
 
 	if (bio->bi_vcnt)
 		return bio_iov_iter_align_down(bio, iter, len_align_mask);
-- 
2.47.3

Re: [PATCH 03/15] block: open code bio_add_page and fix handling of mismatching P2P ranges

[Martin K. Petersen]

Christoph,

> bio_add_page fails to add data to the bio when mixing P2P with non-P2P
> ranges, or ranges that map to different P2P providers.  In that case
> it will trigger that WARN_ON and return an error up the chain instead of
> simply starting a new bio as intended.  Fix this by open coding
> bio_add_page and handling this case explicitly.  While doing so, stop
> merging physical contiguous data that belongs to multiple folios.  While
> this merge could lead to more efficient bio packing in some case,
> dropping will allow to remove handling of this corner case in other
> places and make the code more robust.

Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>

-- 
Martin K. Petersen

[PATCH 04/15] iov_iter: extract a iov_iter_extract_bvecs helper from bio code

[Christoph Hellwig]

Massage __bio_iov_iter_get_pages so that it doesn't need the bio, and
move it to lib/iov_iter.c so that it can be used by block code for
other things than filling a bio and by other subsystems like netfs.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 block/bio.c         | 120 +++++++-------------------------------------
 include/linux/uio.h |   3 ++
 lib/iov_iter.c      |  98 ++++++++++++++++++++++++++++++++++++
 3 files changed, 119 insertions(+), 102 deletions(-)

diff --git a/block/bio.c b/block/bio.c
index 4591f0ba90f5..530082c8cf0c 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -1172,102 +1172,6 @@ void bio_iov_bvec_set(struct bio *bio, const struct iov_iter *iter)
 	bio_set_flag(bio, BIO_CLONED);
 }
 
-static unsigned int get_contig_folio_len(struct page **pages,
-					 unsigned int *num_pages, size_t left,
-					 size_t offset)
-{
-	struct folio *folio = page_folio(pages[0]);
-	size_t contig_sz = min_t(size_t, PAGE_SIZE - offset, left);
-	unsigned int max_pages, i;
-	size_t folio_offset, len;
-
-	folio_offset = PAGE_SIZE * folio_page_idx(folio, pages[0]) + offset;
-	len = min(folio_size(folio) - folio_offset, left);
-
-	/*
-	 * We might COW a single page in the middle of a large folio, so we have
-	 * to check that all pages belong to the same folio.
-	 */
-	left -= contig_sz;
-	max_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
-	for (i = 1; i < max_pages; i++) {
-		size_t next = min_t(size_t, PAGE_SIZE, left);
-
-		if (page_folio(pages[i]) != folio ||
-		    pages[i] != pages[i - 1] + 1)
-			break;
-		contig_sz += next;
-		left -= next;
-	}
-
-	*num_pages = i;
-	return contig_sz;
-}
-
-#define PAGE_PTRS_PER_BVEC     (sizeof(struct bio_vec) / sizeof(struct page *))
-
-/**
- * __bio_iov_iter_get_pages - pin user or kernel pages and add them to a bio
- * @bio: bio to add pages to
- * @iter: iov iterator describing the region to be mapped
- *
- * Extracts pages from *iter and appends them to @bio's bvec array.  The pages
- * will have to be cleaned up in the way indicated by the BIO_PAGE_PINNED flag.
- * For a multi-segment *iter, this function only adds pages from the next
- * non-empty segment of the iov iterator.
- */
-static ssize_t __bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
-{
-	iov_iter_extraction_t extraction_flags = 0;
-	unsigned short nr_pages = bio->bi_max_vecs - bio->bi_vcnt;
-	unsigned short entries_left = bio->bi_max_vecs - bio->bi_vcnt;
-	struct bio_vec *bv = bio->bi_io_vec + bio->bi_vcnt;
-	struct page **pages = (struct page **)bv;
-	ssize_t size;
-	unsigned int i = 0;
-	size_t offset, left, len;
-
-	/*
-	 * Move page array up in the allocated memory for the bio vecs as far as
-	 * possible so that we can start filling biovecs from the beginning
-	 * without overwriting the temporary page array.
-	 */
-	BUILD_BUG_ON(PAGE_PTRS_PER_BVEC < 2);
-	pages += entries_left * (PAGE_PTRS_PER_BVEC - 1);
-
-	if (bio->bi_bdev && blk_queue_pci_p2pdma(bio->bi_bdev->bd_disk->queue))
-		extraction_flags |= ITER_ALLOW_P2PDMA;
-
-	size = iov_iter_extract_pages(iter, &pages,
-				      BIO_MAX_SIZE - bio->bi_iter.bi_size,
-				      nr_pages, extraction_flags, &offset);
-	if (unlikely(size <= 0))
-		return size ? size : -EFAULT;
-
-	nr_pages = DIV_ROUND_UP(offset + size, PAGE_SIZE);
-	for (left = size; left > 0; left -= len) {
-		unsigned int nr_to_add;
-
-		if (bio->bi_vcnt > 0) {
-			struct bio_vec *prev = &bio->bi_io_vec[bio->bi_vcnt - 1];
-
-			if (!zone_device_pages_have_same_pgmap(prev->bv_page,
-					pages[i]))
-				break;
-		}
-
-		len = get_contig_folio_len(&pages[i], &nr_to_add, left, offset);
-		__bio_add_page(bio, pages[i], len, offset);
-		i += nr_to_add;
-		offset = 0;
-	}
-
-	iov_iter_revert(iter, left);
-	while (i < nr_pages)
-		bio_release_page(bio, pages[i++]);
-	return size - left;
-}
-
 /*
  * Aligns the bio size to the len_align_mask, releasing excessive bio vecs that
  * __bio_iov_iter_get_pages may have inserted, and reverts the trimmed length
@@ -1325,7 +1229,7 @@ static int bio_iov_iter_align_down(struct bio *bio, struct iov_iter *iter,
 int bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter,
 			   unsigned len_align_mask)
 {
-	ssize_t ret;
+	iov_iter_extraction_t flags = 0;
 
 	if (WARN_ON_ONCE(bio_flagged(bio, BIO_CLONED)))
 		return -EIO;
@@ -1338,14 +1242,26 @@ int bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter,
 
 	if (iov_iter_extract_will_pin(iter))
 		bio_set_flag(bio, BIO_PAGE_PINNED);
+	if (bio->bi_bdev && blk_queue_pci_p2pdma(bio->bi_bdev->bd_disk->queue))
+		flags |= ITER_ALLOW_P2PDMA;
 
 	do {
-		ret = __bio_iov_iter_get_pages(bio, iter);
-	} while (ret > 0 && iov_iter_count(iter) && !bio_full(bio, 0));
+		ssize_t ret;
+
+		ret = iov_iter_extract_bvecs(iter, bio->bi_io_vec,
+				BIO_MAX_SIZE - bio->bi_iter.bi_size,
+				&bio->bi_vcnt, bio->bi_max_vecs, flags);
+		if (ret <= 0) {
+			if (!bio->bi_vcnt)
+				return ret;
+			break;
+		}
+		bio->bi_iter.bi_size += ret;
+	} while (iov_iter_count(iter) && !bio_full(bio, 0));
 
-	if (bio->bi_vcnt)
-		return bio_iov_iter_align_down(bio, iter, len_align_mask);
-	return ret;
+	if (is_pci_p2pdma_page(bio->bi_io_vec->bv_page))
+		bio->bi_opf |= REQ_NOMERGE;
+	return bio_iov_iter_align_down(bio, iter, len_align_mask);
 }
 
 static void submit_bio_wait_endio(struct bio *bio)
diff --git a/include/linux/uio.h b/include/linux/uio.h
index 5b127043a151..a9bc5b3067e3 100644
--- a/include/linux/uio.h
+++ b/include/linux/uio.h
@@ -389,6 +389,9 @@ ssize_t iov_iter_extract_pages(struct iov_iter *i, struct page ***pages,
 			       size_t maxsize, unsigned int maxpages,
 			       iov_iter_extraction_t extraction_flags,
 			       size_t *offset0);
+ssize_t iov_iter_extract_bvecs(struct iov_iter *iter, struct bio_vec *bv,
+		size_t max_size, unsigned short *nr_vecs,
+		unsigned short max_vecs, iov_iter_extraction_t extraction_flags);
 
 /**
  * iov_iter_extract_will_pin - Indicate how pages from the iterator will be retained
diff --git a/lib/iov_iter.c b/lib/iov_iter.c
index 896760bad455..545250507f08 100644
--- a/lib/iov_iter.c
+++ b/lib/iov_iter.c
@@ -1845,3 +1845,101 @@ ssize_t iov_iter_extract_pages(struct iov_iter *i,
 	return -EFAULT;
 }
 EXPORT_SYMBOL_GPL(iov_iter_extract_pages);
+
+static unsigned int get_contig_folio_len(struct page **pages,
+		unsigned int *num_pages, size_t left, size_t offset)
+{
+	struct folio *folio = page_folio(pages[0]);
+	size_t contig_sz = min_t(size_t, PAGE_SIZE - offset, left);
+	unsigned int max_pages, i;
+	size_t folio_offset, len;
+
+	folio_offset = PAGE_SIZE * folio_page_idx(folio, pages[0]) + offset;
+	len = min(folio_size(folio) - folio_offset, left);
+
+	/*
+	 * We might COW a single page in the middle of a large folio, so we have
+	 * to check that all pages belong to the same folio.
+	 */
+	left -= contig_sz;
+	max_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
+	for (i = 1; i < max_pages; i++) {
+		size_t next = min_t(size_t, PAGE_SIZE, left);
+
+		if (page_folio(pages[i]) != folio ||
+		    pages[i] != pages[i - 1] + 1)
+			break;
+		contig_sz += next;
+		left -= next;
+	}
+
+	*num_pages = i;
+	return contig_sz;
+}
+
+#define PAGE_PTRS_PER_BVEC     (sizeof(struct bio_vec) / sizeof(struct page *))
+
+/**
+ * iov_iter_extract_bvecs - Extract bvecs from an iterator
+ * @iter:	the iterator to extract from
+ * @bv:		bvec return array
+ * @max_size:	maximum size to extract from @iter
+ * @nr_vecs:	number of vectors in @bv (on in and output)
+ * @max_vecs:	maximum vectors in @bv, including those filled before calling
+ * @extraction_flags: flags to qualify request
+ *
+ * Like iov_iter_extract_pages(), but returns physically contiguous ranges
+ * contained in a single folio as a single bvec instead of multiple entries.
+ *
+ * Returns the number of bytes extracted when successful, or a negative errno.
+ * If @nr_vecs was non-zero on entry, the number of successfully extracted bytes
+ * can be 0.
+ */
+ssize_t iov_iter_extract_bvecs(struct iov_iter *iter, struct bio_vec *bv,
+		size_t max_size, unsigned short *nr_vecs,
+		unsigned short max_vecs, iov_iter_extraction_t extraction_flags)
+{
+	unsigned short entries_left = max_vecs - *nr_vecs;
+	unsigned short nr_pages, i = 0;
+	size_t left, offset, len;
+	struct page **pages;
+	ssize_t size;
+
+	/*
+	 * Move page array up in the allocated memory for the bio vecs as far as
+	 * possible so that we can start filling biovecs from the beginning
+	 * without overwriting the temporary page array.
+	 */
+	BUILD_BUG_ON(PAGE_PTRS_PER_BVEC < 2);
+	pages = (struct page **)(bv + *nr_vecs) +
+		entries_left * (PAGE_PTRS_PER_BVEC - 1);
+
+	size = iov_iter_extract_pages(iter, &pages, max_size, entries_left,
+			extraction_flags, &offset);
+	if (unlikely(size <= 0))
+		return size ? size : -EFAULT;
+
+	nr_pages = DIV_ROUND_UP(offset + size, PAGE_SIZE);
+	for (left = size; left > 0; left -= len) {
+		unsigned int nr_to_add;
+
+		if (*nr_vecs > 0 &&
+		    !zone_device_pages_have_same_pgmap(bv[*nr_vecs - 1].bv_page,
+				pages[i]))
+			break;
+
+		len = get_contig_folio_len(&pages[i], &nr_to_add, left, offset);
+		bvec_set_page(&bv[*nr_vecs], pages[i], len, offset);
+		i += nr_to_add;
+		(*nr_vecs)++;
+		offset = 0;
+	}
+
+	iov_iter_revert(iter, left);
+	if (iov_iter_extract_will_pin(iter)) {
+		while (i < nr_pages)
+			unpin_user_page(pages[i++]);
+	}
+	return size - left;
+}
+EXPORT_SYMBOL_GPL(iov_iter_extract_bvecs);
-- 
2.47.3

Re: [PATCH 04/15] iov_iter: extract a iov_iter_extract_bvecs helper from bio code

[Damien Le Moal]

On 1/26/26 2:53 PM, Christoph Hellwig wrote:
> Massage __bio_iov_iter_get_pages so that it doesn't need the bio, and
> move it to lib/iov_iter.c so that it can be used by block code for
> other things than filling a bio and by other subsystems like netfs.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>


> +ssize_t iov_iter_extract_bvecs(struct iov_iter *iter, struct bio_vec *bv,
> +		size_t max_size, unsigned short *nr_vecs,
> +		unsigned short max_vecs, iov_iter_extraction_t extraction_flags)
> +{
> +	unsigned short entries_left = max_vecs - *nr_vecs;

Do we need to check that *nrvecs > 0 && *nrvecs < max_vecs ?
Also, if *nr_vecs == max_vecs, we should warn and return 0, no ?



-- 
Damien Le Moal
Western Digital Research

Re: [PATCH 04/15] iov_iter: extract a iov_iter_extract_bvecs helper from bio code

[Christoph Hellwig]

On Mon, Jan 26, 2026 at 03:27:16PM +0900, Damien Le Moal wrote:
> On 1/26/26 2:53 PM, Christoph Hellwig wrote:
> > Massage __bio_iov_iter_get_pages so that it doesn't need the bio, and
> > move it to lib/iov_iter.c so that it can be used by block code for
> > other things than filling a bio and by other subsystems like netfs.
> > 
> > Signed-off-by: Christoph Hellwig <hch@lst.de>
> 
> 
> > +ssize_t iov_iter_extract_bvecs(struct iov_iter *iter, struct bio_vec *bv,
> > +		size_t max_size, unsigned short *nr_vecs,
> > +		unsigned short max_vecs, iov_iter_extraction_t extraction_flags)
> > +{
> > +	unsigned short entries_left = max_vecs - *nr_vecs;
> 
> Do we need to check that *nrvecs > 0 && *nrvecs < max_vecs ?
> Also, if *nr_vecs == max_vecs, we should warn and return 0, no ?

*nr_vecs = 0 is fine, and in fact the most common case.

We could add a protection for *nrvecs < max_vecs, but this is a very
low-level API, so we should be able to expect some sanity from the
users.  Especially as it will blow up instantly, including with KASAN
splats.

Re: [PATCH 04/15] iov_iter: extract a iov_iter_extract_bvecs helper from bio code

[Martin K. Petersen]

Christoph,

> Massage __bio_iov_iter_get_pages so that it doesn't need the bio, and
> move it to lib/iov_iter.c so that it can be used by block code for
> other things than filling a bio and by other subsystems like netfs.

Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>

-- 
Martin K. Petersen

[PATCH 05/15] block: remove bio_release_page

[Christoph Hellwig]

Merge bio_release_page into the only remaining caller.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Anuj Gupta <anuj20.g@samsung.com>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
---
 block/bio.c |  4 +++-
 block/blk.h | 11 -----------
 2 files changed, 3 insertions(+), 12 deletions(-)

diff --git a/block/bio.c b/block/bio.c
index 530082c8cf0c..285b573ae82f 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -1195,7 +1195,9 @@ static int bio_iov_iter_align_down(struct bio *bio, struct iov_iter *iter,
 			break;
 		}
 
-		bio_release_page(bio, bv->bv_page);
+		if (bio_flagged(bio, BIO_PAGE_PINNED))
+			unpin_user_page(bv->bv_page);
+
 		bio->bi_vcnt--;
 		nbytes -= bv->bv_len;
 	} while (nbytes);
diff --git a/block/blk.h b/block/blk.h
index 980eef1f5690..886238cae5f1 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -595,17 +595,6 @@ void bdev_set_nr_sectors(struct block_device *bdev, sector_t sectors);
 
 struct gendisk *__alloc_disk_node(struct request_queue *q, int node_id,
 		struct lock_class_key *lkclass);
-
-/*
- * Clean up a page appropriately, where the page may be pinned, may have a
- * ref taken on it or neither.
- */
-static inline void bio_release_page(struct bio *bio, struct page *page)
-{
-	if (bio_flagged(bio, BIO_PAGE_PINNED))
-		unpin_user_page(page);
-}
-
 struct request_queue *blk_alloc_queue(struct queue_limits *lim, int node_id);
 
 int disk_scan_partitions(struct gendisk *disk, blk_mode_t mode);
-- 
2.47.3

Re: [PATCH 05/15] block: remove bio_release_page

[Martin K. Petersen]

Christoph,

> Merge bio_release_page into the only remaining caller.

Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>

-- 
Martin K. Petersen

[PATCH 06/15] block: add helpers to bounce buffer an iov_iter into bios

[Christoph Hellwig]

Add helpers to implement bounce buffering of data into a bio to implement
direct I/O for cases where direct user access is not possible because
stable in-flight data is required.  These are intended to be used as
easily as bio_iov_iter_get_pages for the zero-copy path.

The write side is trivial and just copies data into the bounce buffer.
The read side is a lot more complex because it needs to perform the copy
from the completion context, and without preserving the iov_iter through
the call chain.  It steals a trick from the integrity data user interface
and uses the first vector in the bio for the bounce buffer data that is
fed to the block I/O stack, and uses the others to record the user
buffer fragments.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Anuj Gupta <anuj20.g@samsung.com
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
---
 block/bio.c         | 179 ++++++++++++++++++++++++++++++++++++++++++++
 include/linux/bio.h |  26 +++++++
 2 files changed, 205 insertions(+)

diff --git a/block/bio.c b/block/bio.c
index 285b573ae82f..49f7548a31d6 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -1266,6 +1266,185 @@ int bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter,
 	return bio_iov_iter_align_down(bio, iter, len_align_mask);
 }
 
+static struct folio *folio_alloc_greedy(gfp_t gfp, size_t *size)
+{
+	struct folio *folio;
+
+	while (*size > PAGE_SIZE) {
+		folio = folio_alloc(gfp | __GFP_NORETRY, get_order(*size));
+		if (folio)
+			return folio;
+		*size = rounddown_pow_of_two(*size - 1);
+	}
+
+	return folio_alloc(gfp, get_order(*size));
+}
+
+static void bio_free_folios(struct bio *bio)
+{
+	struct bio_vec *bv;
+	int i;
+
+	bio_for_each_bvec_all(bv, bio, i) {
+		struct folio *folio = page_folio(bv->bv_page);
+
+		if (!is_zero_folio(folio))
+			folio_put(folio);
+	}
+}
+
+static int bio_iov_iter_bounce_write(struct bio *bio, struct iov_iter *iter)
+{
+	size_t total_len = iov_iter_count(iter);
+
+	if (WARN_ON_ONCE(bio_flagged(bio, BIO_CLONED)))
+		return -EINVAL;
+	if (WARN_ON_ONCE(bio->bi_iter.bi_size))
+		return -EINVAL;
+	if (WARN_ON_ONCE(bio->bi_vcnt >= bio->bi_max_vecs))
+		return -EINVAL;
+
+	do {
+		size_t this_len = min(total_len, SZ_1M);
+		struct folio *folio;
+
+		if (this_len > PAGE_SIZE * 2)
+			this_len = rounddown_pow_of_two(this_len);
+
+		if (bio->bi_iter.bi_size > BIO_MAX_SIZE - this_len)
+			break;
+
+		folio = folio_alloc_greedy(GFP_KERNEL, &this_len);
+		if (!folio)
+			break;
+		bio_add_folio_nofail(bio, folio, this_len, 0);
+
+		if (copy_from_iter(folio_address(folio), this_len, iter) !=
+				this_len) {
+			bio_free_folios(bio);
+			return -EFAULT;
+		}
+
+		total_len -= this_len;
+	} while (total_len && bio->bi_vcnt < bio->bi_max_vecs);
+
+	if (!bio->bi_iter.bi_size)
+		return -ENOMEM;
+	return 0;
+}
+
+static int bio_iov_iter_bounce_read(struct bio *bio, struct iov_iter *iter)
+{
+	size_t len = min(iov_iter_count(iter), SZ_1M);
+	struct folio *folio;
+
+	folio = folio_alloc_greedy(GFP_KERNEL, &len);
+	if (!folio)
+		return -ENOMEM;
+
+	do {
+		ssize_t ret;
+
+		ret = iov_iter_extract_bvecs(iter, bio->bi_io_vec + 1, len,
+				&bio->bi_vcnt, bio->bi_max_vecs - 1, 0);
+		if (ret <= 0) {
+			if (!bio->bi_vcnt)
+				return ret;
+			break;
+		}
+		len -= ret;
+		bio->bi_iter.bi_size += ret;
+	} while (len && bio->bi_vcnt < bio->bi_max_vecs - 1);
+
+	/*
+	 * Set the folio directly here.  The above loop has already calculated
+	 * the correct bi_size, and we use bi_vcnt for the user buffers.  That
+	 * is safe as bi_vcnt is only used by the submitter and not the actual
+	 * I/O path.
+	 */
+	bvec_set_folio(&bio->bi_io_vec[0], folio, bio->bi_iter.bi_size, 0);
+	if (iov_iter_extract_will_pin(iter))
+		bio_set_flag(bio, BIO_PAGE_PINNED);
+	return 0;
+}
+
+/**
+ * bio_iov_iter_bounce - bounce buffer data from an iter into a bio
+ * @bio:	bio to send
+ * @iter:	iter to read from / write into
+ *
+ * Helper for direct I/O implementations that need to bounce buffer because
+ * we need to checksum the data or perform other operations that require
+ * consistency.  Allocates folios to back the bounce buffer, and for writes
+ * copies the data into it.  Needs to be paired with bio_iov_iter_unbounce()
+ * called on completion.
+ */
+int bio_iov_iter_bounce(struct bio *bio, struct iov_iter *iter)
+{
+	if (op_is_write(bio_op(bio)))
+		return bio_iov_iter_bounce_write(bio, iter);
+	return bio_iov_iter_bounce_read(bio, iter);
+}
+
+static void bvec_unpin(struct bio_vec *bv, bool mark_dirty)
+{
+	struct folio *folio = page_folio(bv->bv_page);
+	size_t nr_pages = (bv->bv_offset + bv->bv_len - 1) / PAGE_SIZE -
+			bv->bv_offset / PAGE_SIZE + 1;
+
+	if (mark_dirty)
+		folio_mark_dirty_lock(folio);
+	unpin_user_folio(folio, nr_pages);
+}
+
+static void bio_iov_iter_unbounce_read(struct bio *bio, bool is_error,
+		bool mark_dirty)
+{
+	unsigned int len = bio->bi_io_vec[0].bv_len;
+
+	if (likely(!is_error)) {
+		void *buf = bvec_virt(&bio->bi_io_vec[0]);
+		struct iov_iter to;
+
+		iov_iter_bvec(&to, ITER_DEST, bio->bi_io_vec + 1, bio->bi_vcnt,
+				len);
+		/* copying to pinned pages should always work */
+		WARN_ON_ONCE(copy_to_iter(buf, len, &to) != len);
+	} else {
+		/* No need to mark folios dirty if never copied to them */
+		mark_dirty = false;
+	}
+
+	if (bio_flagged(bio, BIO_PAGE_PINNED)) {
+		int i;
+
+		for (i = 0; i < bio->bi_vcnt; i++)
+			bvec_unpin(&bio->bi_io_vec[1 + i], mark_dirty);
+	}
+
+	folio_put(page_folio(bio->bi_io_vec[0].bv_page));
+}
+
+/**
+ * bio_iov_iter_unbounce - finish a bounce buffer operation
+ * @bio:	completed bio
+ * @is_error:	%true if an I/O error occurred and data should not be copied
+ * @mark_dirty:	If %true, folios will be marked dirty.
+ *
+ * Helper for direct I/O implementations that need to bounce buffer because
+ * we need to checksum the data or perform other operations that require
+ * consistency.  Called to complete a bio set up by bio_iov_iter_bounce().
+ * Copies data back for reads, and marks the original folios dirty if
+ * requested and then frees the bounce buffer.
+ */
+void bio_iov_iter_unbounce(struct bio *bio, bool is_error, bool mark_dirty)
+{
+	if (op_is_write(bio_op(bio)))
+		bio_free_folios(bio);
+	else
+		bio_iov_iter_unbounce_read(bio, is_error, mark_dirty);
+}
+
 static void submit_bio_wait_endio(struct bio *bio)
 {
 	complete(bio->bi_private);
diff --git a/include/linux/bio.h b/include/linux/bio.h
index c75a9b3672aa..95cfc79b88b8 100644
--- a/include/linux/bio.h
+++ b/include/linux/bio.h
@@ -403,6 +403,29 @@ static inline int bio_iov_vecs_to_alloc(struct iov_iter *iter, int max_segs)
 	return iov_iter_npages(iter, max_segs);
 }
 
+/**
+ * bio_iov_bounce_nr_vecs - calculate number of bvecs for a bounce bio
+ * @iter:	iter to bounce from
+ * @op:		REQ_OP_* for the bio
+ *
+ * Calculates how many bvecs are needed for the next bio to bounce from/to
+ * @iter.
+ */
+static inline unsigned short
+bio_iov_bounce_nr_vecs(struct iov_iter *iter, blk_opf_t op)
+{
+	/*
+	 * We still need to bounce bvec iters, so don't special case them
+	 * here unlike in bio_iov_vecs_to_alloc.
+	 *
+	 * For reads we need to use a vector for the bounce buffer, account
+	 * for that here.
+	 */
+	if (op_is_write(op))
+		return iov_iter_npages(iter, BIO_MAX_VECS);
+	return iov_iter_npages(iter, BIO_MAX_VECS - 1) + 1;
+}
+
 struct request_queue;
 
 void bio_init(struct bio *bio, struct block_device *bdev, struct bio_vec *table,
@@ -456,6 +479,9 @@ void __bio_release_pages(struct bio *bio, bool mark_dirty);
 extern void bio_set_pages_dirty(struct bio *bio);
 extern void bio_check_pages_dirty(struct bio *bio);
 
+int bio_iov_iter_bounce(struct bio *bio, struct iov_iter *iter);
+void bio_iov_iter_unbounce(struct bio *bio, bool is_error, bool mark_dirty);
+
 extern void bio_copy_data_iter(struct bio *dst, struct bvec_iter *dst_iter,
 			       struct bio *src, struct bvec_iter *src_iter);
 extern void bio_copy_data(struct bio *dst, struct bio *src);
-- 
2.47.3

Re: [PATCH 06/15] block: add helpers to bounce buffer an iov_iter into bios

[Darrick J. Wong]

On Mon, Jan 26, 2026 at 06:53:37AM +0100, Christoph Hellwig wrote:
> Add helpers to implement bounce buffering of data into a bio to implement
> direct I/O for cases where direct user access is not possible because
> stable in-flight data is required.  These are intended to be used as
> easily as bio_iov_iter_get_pages for the zero-copy path.
> 
> The write side is trivial and just copies data into the bounce buffer.
> The read side is a lot more complex because it needs to perform the copy
> from the completion context, and without preserving the iov_iter through
> the call chain.  It steals a trick from the integrity data user interface
> and uses the first vector in the bio for the bounce buffer data that is
> fed to the block I/O stack, and uses the others to record the user
> buffer fragments.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> Reviewed-by: Anuj Gupta <anuj20.g@samsung.com
> Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
> Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
> Tested-by: Anuj Gupta <anuj20.g@samsung.com>

Looks good, thanks for addressing my feedback on the last revision :)
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>

--D

> ---
>  block/bio.c         | 179 ++++++++++++++++++++++++++++++++++++++++++++
>  include/linux/bio.h |  26 +++++++
>  2 files changed, 205 insertions(+)
> 
> diff --git a/block/bio.c b/block/bio.c
> index 285b573ae82f..49f7548a31d6 100644
> --- a/block/bio.c
> +++ b/block/bio.c
> @@ -1266,6 +1266,185 @@ int bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter,
>  	return bio_iov_iter_align_down(bio, iter, len_align_mask);
>  }
>  
> +static struct folio *folio_alloc_greedy(gfp_t gfp, size_t *size)
> +{
> +	struct folio *folio;
> +
> +	while (*size > PAGE_SIZE) {
> +		folio = folio_alloc(gfp | __GFP_NORETRY, get_order(*size));
> +		if (folio)
> +			return folio;
> +		*size = rounddown_pow_of_two(*size - 1);
> +	}
> +
> +	return folio_alloc(gfp, get_order(*size));
> +}
> +
> +static void bio_free_folios(struct bio *bio)
> +{
> +	struct bio_vec *bv;
> +	int i;
> +
> +	bio_for_each_bvec_all(bv, bio, i) {
> +		struct folio *folio = page_folio(bv->bv_page);
> +
> +		if (!is_zero_folio(folio))
> +			folio_put(folio);
> +	}
> +}
> +
> +static int bio_iov_iter_bounce_write(struct bio *bio, struct iov_iter *iter)
> +{
> +	size_t total_len = iov_iter_count(iter);
> +
> +	if (WARN_ON_ONCE(bio_flagged(bio, BIO_CLONED)))
> +		return -EINVAL;
> +	if (WARN_ON_ONCE(bio->bi_iter.bi_size))
> +		return -EINVAL;
> +	if (WARN_ON_ONCE(bio->bi_vcnt >= bio->bi_max_vecs))
> +		return -EINVAL;
> +
> +	do {
> +		size_t this_len = min(total_len, SZ_1M);
> +		struct folio *folio;
> +
> +		if (this_len > PAGE_SIZE * 2)
> +			this_len = rounddown_pow_of_two(this_len);
> +
> +		if (bio->bi_iter.bi_size > BIO_MAX_SIZE - this_len)
> +			break;
> +
> +		folio = folio_alloc_greedy(GFP_KERNEL, &this_len);
> +		if (!folio)
> +			break;
> +		bio_add_folio_nofail(bio, folio, this_len, 0);
> +
> +		if (copy_from_iter(folio_address(folio), this_len, iter) !=
> +				this_len) {
> +			bio_free_folios(bio);
> +			return -EFAULT;
> +		}
> +
> +		total_len -= this_len;
> +	} while (total_len && bio->bi_vcnt < bio->bi_max_vecs);
> +
> +	if (!bio->bi_iter.bi_size)
> +		return -ENOMEM;
> +	return 0;
> +}
> +
> +static int bio_iov_iter_bounce_read(struct bio *bio, struct iov_iter *iter)
> +{
> +	size_t len = min(iov_iter_count(iter), SZ_1M);
> +	struct folio *folio;
> +
> +	folio = folio_alloc_greedy(GFP_KERNEL, &len);
> +	if (!folio)
> +		return -ENOMEM;
> +
> +	do {
> +		ssize_t ret;
> +
> +		ret = iov_iter_extract_bvecs(iter, bio->bi_io_vec + 1, len,
> +				&bio->bi_vcnt, bio->bi_max_vecs - 1, 0);
> +		if (ret <= 0) {
> +			if (!bio->bi_vcnt)
> +				return ret;
> +			break;
> +		}
> +		len -= ret;
> +		bio->bi_iter.bi_size += ret;
> +	} while (len && bio->bi_vcnt < bio->bi_max_vecs - 1);
> +
> +	/*
> +	 * Set the folio directly here.  The above loop has already calculated
> +	 * the correct bi_size, and we use bi_vcnt for the user buffers.  That
> +	 * is safe as bi_vcnt is only used by the submitter and not the actual
> +	 * I/O path.
> +	 */
> +	bvec_set_folio(&bio->bi_io_vec[0], folio, bio->bi_iter.bi_size, 0);
> +	if (iov_iter_extract_will_pin(iter))
> +		bio_set_flag(bio, BIO_PAGE_PINNED);
> +	return 0;
> +}
> +
> +/**
> + * bio_iov_iter_bounce - bounce buffer data from an iter into a bio
> + * @bio:	bio to send
> + * @iter:	iter to read from / write into
> + *
> + * Helper for direct I/O implementations that need to bounce buffer because
> + * we need to checksum the data or perform other operations that require
> + * consistency.  Allocates folios to back the bounce buffer, and for writes
> + * copies the data into it.  Needs to be paired with bio_iov_iter_unbounce()
> + * called on completion.
> + */
> +int bio_iov_iter_bounce(struct bio *bio, struct iov_iter *iter)
> +{
> +	if (op_is_write(bio_op(bio)))
> +		return bio_iov_iter_bounce_write(bio, iter);
> +	return bio_iov_iter_bounce_read(bio, iter);
> +}
> +
> +static void bvec_unpin(struct bio_vec *bv, bool mark_dirty)
> +{
> +	struct folio *folio = page_folio(bv->bv_page);
> +	size_t nr_pages = (bv->bv_offset + bv->bv_len - 1) / PAGE_SIZE -
> +			bv->bv_offset / PAGE_SIZE + 1;
> +
> +	if (mark_dirty)
> +		folio_mark_dirty_lock(folio);
> +	unpin_user_folio(folio, nr_pages);
> +}
> +
> +static void bio_iov_iter_unbounce_read(struct bio *bio, bool is_error,
> +		bool mark_dirty)
> +{
> +	unsigned int len = bio->bi_io_vec[0].bv_len;
> +
> +	if (likely(!is_error)) {
> +		void *buf = bvec_virt(&bio->bi_io_vec[0]);
> +		struct iov_iter to;
> +
> +		iov_iter_bvec(&to, ITER_DEST, bio->bi_io_vec + 1, bio->bi_vcnt,
> +				len);
> +		/* copying to pinned pages should always work */
> +		WARN_ON_ONCE(copy_to_iter(buf, len, &to) != len);
> +	} else {
> +		/* No need to mark folios dirty if never copied to them */
> +		mark_dirty = false;
> +	}
> +
> +	if (bio_flagged(bio, BIO_PAGE_PINNED)) {
> +		int i;
> +
> +		for (i = 0; i < bio->bi_vcnt; i++)
> +			bvec_unpin(&bio->bi_io_vec[1 + i], mark_dirty);
> +	}
> +
> +	folio_put(page_folio(bio->bi_io_vec[0].bv_page));
> +}
> +
> +/**
> + * bio_iov_iter_unbounce - finish a bounce buffer operation
> + * @bio:	completed bio
> + * @is_error:	%true if an I/O error occurred and data should not be copied
> + * @mark_dirty:	If %true, folios will be marked dirty.
> + *
> + * Helper for direct I/O implementations that need to bounce buffer because
> + * we need to checksum the data or perform other operations that require
> + * consistency.  Called to complete a bio set up by bio_iov_iter_bounce().
> + * Copies data back for reads, and marks the original folios dirty if
> + * requested and then frees the bounce buffer.
> + */
> +void bio_iov_iter_unbounce(struct bio *bio, bool is_error, bool mark_dirty)
> +{
> +	if (op_is_write(bio_op(bio)))
> +		bio_free_folios(bio);
> +	else
> +		bio_iov_iter_unbounce_read(bio, is_error, mark_dirty);
> +}
> +
>  static void submit_bio_wait_endio(struct bio *bio)
>  {
>  	complete(bio->bi_private);
> diff --git a/include/linux/bio.h b/include/linux/bio.h
> index c75a9b3672aa..95cfc79b88b8 100644
> --- a/include/linux/bio.h
> +++ b/include/linux/bio.h
> @@ -403,6 +403,29 @@ static inline int bio_iov_vecs_to_alloc(struct iov_iter *iter, int max_segs)
>  	return iov_iter_npages(iter, max_segs);
>  }
>  
> +/**
> + * bio_iov_bounce_nr_vecs - calculate number of bvecs for a bounce bio
> + * @iter:	iter to bounce from
> + * @op:		REQ_OP_* for the bio
> + *
> + * Calculates how many bvecs are needed for the next bio to bounce from/to
> + * @iter.
> + */
> +static inline unsigned short
> +bio_iov_bounce_nr_vecs(struct iov_iter *iter, blk_opf_t op)
> +{
> +	/*
> +	 * We still need to bounce bvec iters, so don't special case them
> +	 * here unlike in bio_iov_vecs_to_alloc.
> +	 *
> +	 * For reads we need to use a vector for the bounce buffer, account
> +	 * for that here.
> +	 */
> +	if (op_is_write(op))
> +		return iov_iter_npages(iter, BIO_MAX_VECS);
> +	return iov_iter_npages(iter, BIO_MAX_VECS - 1) + 1;
> +}
> +
>  struct request_queue;
>  
>  void bio_init(struct bio *bio, struct block_device *bdev, struct bio_vec *table,
> @@ -456,6 +479,9 @@ void __bio_release_pages(struct bio *bio, bool mark_dirty);
>  extern void bio_set_pages_dirty(struct bio *bio);
>  extern void bio_check_pages_dirty(struct bio *bio);
>  
> +int bio_iov_iter_bounce(struct bio *bio, struct iov_iter *iter);
> +void bio_iov_iter_unbounce(struct bio *bio, bool is_error, bool mark_dirty);
> +
>  extern void bio_copy_data_iter(struct bio *dst, struct bvec_iter *dst_iter,
>  			       struct bio *src, struct bvec_iter *src_iter);
>  extern void bio_copy_data(struct bio *dst, struct bio *src);
> -- 
> 2.47.3
> 
> 

Re: [PATCH 06/15] block: add helpers to bounce buffer an iov_iter into bios

[Martin K. Petersen]

Christoph,

> Add helpers to implement bounce buffering of data into a bio to
> implement direct I/O for cases where direct user access is not
> possible because stable in-flight data is required. These are intended
> to be used as easily as bio_iov_iter_get_pages for the zero-copy path.

Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>

-- 
Martin K. Petersen

[PATCH 07/15] iomap: fix submission side handling of completion side errors

[Christoph Hellwig]

The "if (dio->error)" in iomap_dio_bio_iter exists to stop submitting
more bios when a completion already return an error.  Commit cfe057f7db1f
("iomap_dio_actor(): fix iov_iter bugs") made it revert the iov by
"copied", which is very wrong given that we've already consumed that
range and submitted a bio for it.

Fixes: cfe057f7db1f ("iomap_dio_actor(): fix iov_iter bugs")
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
---
 fs/iomap/direct-io.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index 4000c8596d9b..867c0ac6df8f 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -443,9 +443,13 @@ static int iomap_dio_bio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
 	nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS);
 	do {
 		size_t n;
-		if (dio->error) {
-			iov_iter_revert(dio->submit.iter, copied);
-			copied = ret = 0;
+
+		/*
+		 * If completions already occurred and reported errors, give up now and
+		 * don't bother submitting more bios.
+		 */
+		if (unlikely(data_race(dio->error))) {
+			ret = 0;
 			goto out;
 		}
 
-- 
2.47.3

[PATCH 08/15] iomap: simplify iomap_dio_bio_iter

[Christoph Hellwig]

Use iov_iter_count to check if we need to continue as that just reads
a field in the iov_iter, and only use bio_iov_vecs_to_alloc to calculate
the actual number of vectors to allocate for the bio.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
---
 fs/iomap/direct-io.c | 13 ++++++-------
 1 file changed, 6 insertions(+), 7 deletions(-)

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index 867c0ac6df8f..de03bc7cf4ed 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -312,7 +312,7 @@ static int iomap_dio_bio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
 	blk_opf_t bio_opf = REQ_SYNC | REQ_IDLE;
 	struct bio *bio;
 	bool need_zeroout = false;
-	int nr_pages, ret = 0;
+	int ret = 0;
 	u64 copied = 0;
 	size_t orig_count;
 	unsigned int alignment;
@@ -440,7 +440,6 @@ static int iomap_dio_bio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
 			goto out;
 	}
 
-	nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS);
 	do {
 		size_t n;
 
@@ -453,7 +452,9 @@ static int iomap_dio_bio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
 			goto out;
 		}
 
-		bio = iomap_dio_alloc_bio(iter, dio, nr_pages, bio_opf);
+		bio = iomap_dio_alloc_bio(iter, dio,
+				bio_iov_vecs_to_alloc(dio->submit.iter,
+						BIO_MAX_VECS), bio_opf);
 		fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
 					  GFP_KERNEL);
 		bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
@@ -495,16 +496,14 @@ static int iomap_dio_bio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
 		dio->size += n;
 		copied += n;
 
-		nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter,
-						 BIO_MAX_VECS);
 		/*
 		 * We can only poll for single bio I/Os.
 		 */
-		if (nr_pages)
+		if (iov_iter_count(dio->submit.iter))
 			dio->iocb->ki_flags &= ~IOCB_HIPRI;
 		iomap_dio_submit_bio(iter, dio, bio, pos);
 		pos += n;
-	} while (nr_pages);
+	} while (iov_iter_count(dio->submit.iter));
 
 	/*
 	 * We need to zeroout the tail of a sub-block write if the extent type
-- 
2.47.3

[PATCH 09/15] iomap: split out the per-bio logic from iomap_dio_bio_iter

[Christoph Hellwig]

Factor out a separate helper that builds and submits a single bio.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
---
 fs/iomap/direct-io.c | 111 +++++++++++++++++++++++--------------------
 1 file changed, 59 insertions(+), 52 deletions(-)

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index de03bc7cf4ed..bb79519dec65 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -302,6 +302,56 @@ static int iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio,
 	return 0;
 }
 
+static ssize_t iomap_dio_bio_iter_one(struct iomap_iter *iter,
+		struct iomap_dio *dio, loff_t pos, unsigned int alignment,
+		blk_opf_t op)
+{
+	struct bio *bio;
+	ssize_t ret;
+
+	bio = iomap_dio_alloc_bio(iter, dio,
+			bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS),
+			op);
+	fscrypt_set_bio_crypt_ctx(bio, iter->inode,
+			pos >> iter->inode->i_blkbits, GFP_KERNEL);
+	bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos);
+	bio->bi_write_hint = iter->inode->i_write_hint;
+	bio->bi_ioprio = dio->iocb->ki_ioprio;
+	bio->bi_private = dio;
+	bio->bi_end_io = iomap_dio_bio_end_io;
+
+	ret = bio_iov_iter_get_pages(bio, dio->submit.iter, alignment - 1);
+	if (unlikely(ret))
+		goto out_put_bio;
+	ret = bio->bi_iter.bi_size;
+
+	/*
+	 * An atomic write bio must cover the complete length.  If it doesn't,
+	 * error out.
+	 */
+	if ((op & REQ_ATOMIC) && WARN_ON_ONCE(ret != iomap_length(iter))) {
+		ret = -EINVAL;
+		goto out_put_bio;
+	}
+
+	if (dio->flags & IOMAP_DIO_WRITE)
+		task_io_account_write(ret);
+	else if (dio->flags & IOMAP_DIO_DIRTY)
+		bio_set_pages_dirty(bio);
+
+	/*
+	 * We can only poll for single bio I/Os.
+	 */
+	if (iov_iter_count(dio->submit.iter))
+		dio->iocb->ki_flags &= ~IOCB_HIPRI;
+	iomap_dio_submit_bio(iter, dio, bio, pos);
+	return ret;
+
+out_put_bio:
+	bio_put(bio);
+	return ret;
+}
+
 static int iomap_dio_bio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
 {
 	const struct iomap *iomap = &iter->iomap;
@@ -310,12 +360,11 @@ static int iomap_dio_bio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
 	const loff_t length = iomap_length(iter);
 	loff_t pos = iter->pos;
 	blk_opf_t bio_opf = REQ_SYNC | REQ_IDLE;
-	struct bio *bio;
 	bool need_zeroout = false;
-	int ret = 0;
 	u64 copied = 0;
 	size_t orig_count;
 	unsigned int alignment;
+	ssize_t ret = 0;
 
 	/*
 	 * File systems that write out of place and always allocate new blocks
@@ -441,68 +490,27 @@ static int iomap_dio_bio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
 	}
 
 	do {
-		size_t n;
-
 		/*
 		 * If completions already occurred and reported errors, give up now and
 		 * don't bother submitting more bios.
 		 */
-		if (unlikely(data_race(dio->error))) {
-			ret = 0;
+		if (unlikely(data_race(dio->error)))
 			goto out;
-		}
 
-		bio = iomap_dio_alloc_bio(iter, dio,
-				bio_iov_vecs_to_alloc(dio->submit.iter,
-						BIO_MAX_VECS), bio_opf);
-		fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
-					  GFP_KERNEL);
-		bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
-		bio->bi_write_hint = inode->i_write_hint;
-		bio->bi_ioprio = dio->iocb->ki_ioprio;
-		bio->bi_private = dio;
-		bio->bi_end_io = iomap_dio_bio_end_io;
-
-		ret = bio_iov_iter_get_pages(bio, dio->submit.iter,
-					     alignment - 1);
-		if (unlikely(ret)) {
+		ret = iomap_dio_bio_iter_one(iter, dio, pos, alignment, bio_opf);
+		if (unlikely(ret < 0)) {
 			/*
 			 * We have to stop part way through an IO. We must fall
 			 * through to the sub-block tail zeroing here, otherwise
 			 * this short IO may expose stale data in the tail of
 			 * the block we haven't written data to.
 			 */
-			bio_put(bio);
-			goto zero_tail;
-		}
-
-		n = bio->bi_iter.bi_size;
-		if (WARN_ON_ONCE((bio_opf & REQ_ATOMIC) && n != length)) {
-			/*
-			 * An atomic write bio must cover the complete length,
-			 * which it doesn't, so error. We may need to zero out
-			 * the tail (complete FS block), similar to when
-			 * bio_iov_iter_get_pages() returns an error, above.
-			 */
-			ret = -EINVAL;
-			bio_put(bio);
-			goto zero_tail;
+			break;
 		}
-		if (dio->flags & IOMAP_DIO_WRITE)
-			task_io_account_write(n);
-		else if (dio->flags & IOMAP_DIO_DIRTY)
-			bio_set_pages_dirty(bio);
-
-		dio->size += n;
-		copied += n;
-
-		/*
-		 * We can only poll for single bio I/Os.
-		 */
-		if (iov_iter_count(dio->submit.iter))
-			dio->iocb->ki_flags &= ~IOCB_HIPRI;
-		iomap_dio_submit_bio(iter, dio, bio, pos);
-		pos += n;
+		dio->size += ret;
+		copied += ret;
+		pos += ret;
+		ret = 0;
 	} while (iov_iter_count(dio->submit.iter));
 
 	/*
@@ -511,7 +519,6 @@ static int iomap_dio_bio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
 	 * the block tail in the latter case, we can expose stale data via mmap
 	 * reads of the EOF block.
 	 */
-zero_tail:
 	if (need_zeroout ||
 	    ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
 		/* zero out from the end of the write to the end of the block */
-- 
2.47.3

[PATCH 10/15] iomap: share code between iomap_dio_bio_end_io and iomap_finish_ioend_direct

[Christoph Hellwig]

Refactor the two per-bio completion handlers to share common code using
a new helper.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
---
 fs/iomap/direct-io.c | 42 +++++++++++++++++++-----------------------
 1 file changed, 19 insertions(+), 23 deletions(-)

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index bb79519dec65..c1d5db85c8c7 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -211,16 +211,20 @@ static void iomap_dio_done(struct iomap_dio *dio)
 	iomap_dio_complete_work(&dio->aio.work);
 }
 
-void iomap_dio_bio_end_io(struct bio *bio)
+static void __iomap_dio_bio_end_io(struct bio *bio, bool inline_completion)
 {
 	struct iomap_dio *dio = bio->bi_private;
 	bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
 
-	if (bio->bi_status)
-		iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
-
-	if (atomic_dec_and_test(&dio->ref))
+	if (atomic_dec_and_test(&dio->ref)) {
+		/*
+		 * Avoid another context switch for the completion when already
+		 * called from the ioend completion workqueue.
+		 */
+		if (inline_completion)
+			dio->flags &= ~IOMAP_DIO_COMP_WORK;
 		iomap_dio_done(dio);
+	}
 
 	if (should_dirty) {
 		bio_check_pages_dirty(bio);
@@ -229,33 +233,25 @@ void iomap_dio_bio_end_io(struct bio *bio)
 		bio_put(bio);
 	}
 }
+
+void iomap_dio_bio_end_io(struct bio *bio)
+{
+	struct iomap_dio *dio = bio->bi_private;
+
+	if (bio->bi_status)
+		iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
+	__iomap_dio_bio_end_io(bio, false);
+}
 EXPORT_SYMBOL_GPL(iomap_dio_bio_end_io);
 
 u32 iomap_finish_ioend_direct(struct iomap_ioend *ioend)
 {
 	struct iomap_dio *dio = ioend->io_bio.bi_private;
-	bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
 	u32 vec_count = ioend->io_bio.bi_vcnt;
 
 	if (ioend->io_error)
 		iomap_dio_set_error(dio, ioend->io_error);
-
-	if (atomic_dec_and_test(&dio->ref)) {
-		/*
-		 * Try to avoid another context switch for the completion given
-		 * that we are already called from the ioend completion
-		 * workqueue.
-		 */
-		dio->flags &= ~IOMAP_DIO_COMP_WORK;
-		iomap_dio_done(dio);
-	}
-
-	if (should_dirty) {
-		bio_check_pages_dirty(&ioend->io_bio);
-	} else {
-		bio_release_pages(&ioend->io_bio, false);
-		bio_put(&ioend->io_bio);
-	}
+	__iomap_dio_bio_end_io(&ioend->io_bio, true);
 
 	/*
 	 * Return the number of bvecs completed as even direct I/O completions
-- 
2.47.3

[PATCH 11/15] iomap: free the bio before completing the dio

[Christoph Hellwig]

There are good arguments for processing the user completions ASAP vs.
freeing resources ASAP, but freeing the bio first here removes potential
use after free hazards when checking flags, and will simplify the
upcoming bounce buffer support.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
---
 fs/iomap/direct-io.c | 17 +++++++++--------
 1 file changed, 9 insertions(+), 8 deletions(-)

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index c1d5db85c8c7..d4d52775ce25 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -214,7 +214,15 @@ static void iomap_dio_done(struct iomap_dio *dio)
 static void __iomap_dio_bio_end_io(struct bio *bio, bool inline_completion)
 {
 	struct iomap_dio *dio = bio->bi_private;
-	bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
+
+	if (dio->flags & IOMAP_DIO_DIRTY) {
+		bio_check_pages_dirty(bio);
+	} else {
+		bio_release_pages(bio, false);
+		bio_put(bio);
+	}
+
+	/* Do not touch bio below, we just gave up our reference. */
 
 	if (atomic_dec_and_test(&dio->ref)) {
 		/*
@@ -225,13 +233,6 @@ static void __iomap_dio_bio_end_io(struct bio *bio, bool inline_completion)
 			dio->flags &= ~IOMAP_DIO_COMP_WORK;
 		iomap_dio_done(dio);
 	}
-
-	if (should_dirty) {
-		bio_check_pages_dirty(bio);
-	} else {
-		bio_release_pages(bio, false);
-		bio_put(bio);
-	}
 }
 
 void iomap_dio_bio_end_io(struct bio *bio)
-- 
2.47.3

Re: [PATCH 11/15] iomap: free the bio before completing the dio

[Damien Le Moal]

On 1/26/26 2:53 PM, Christoph Hellwig wrote:
> There are good arguments for processing the user completions ASAP vs.
> freeing resources ASAP, but freeing the bio first here removes potential
> use after free hazards when checking flags, and will simplify the
> upcoming bounce buffer support.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
> Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>

Repeated tag...

Looks good to me.

Reviewed-by: Damien Le Moal <dlemoal@kernel.org>

(which I think I already sent :))


-- 
Damien Le Moal
Western Digital Research

Re: [PATCH 11/15] iomap: free the bio before completing the dio

[Christoph Hellwig]

On Mon, Jan 26, 2026 at 03:22:41PM +0900, Damien Le Moal wrote:
> > Signed-off-by: Christoph Hellwig <hch@lst.de>
> > Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
> > Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
> 
> Repeated tag...
> 
> Looks good to me.
> 
> Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
> 
> (which I think I already sent :))

I guess I messed up and pasted Darricks' review again instead of yours,
sorry.

[PATCH 12/15] iomap: rename IOMAP_DIO_DIRTY to IOMAP_DIO_USER_BACKED

[Christoph Hellwig]

Match the more descriptive iov_iter terminology instead of encoding
what we do with them for reads only.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
---
 fs/iomap/direct-io.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index d4d52775ce25..eca7adda595a 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -22,7 +22,7 @@
 #define IOMAP_DIO_WRITE_THROUGH	(1U << 28)
 #define IOMAP_DIO_NEED_SYNC	(1U << 29)
 #define IOMAP_DIO_WRITE		(1U << 30)
-#define IOMAP_DIO_DIRTY		(1U << 31)
+#define IOMAP_DIO_USER_BACKED	(1U << 31)
 
 struct iomap_dio {
 	struct kiocb		*iocb;
@@ -215,7 +215,7 @@ static void __iomap_dio_bio_end_io(struct bio *bio, bool inline_completion)
 {
 	struct iomap_dio *dio = bio->bi_private;
 
-	if (dio->flags & IOMAP_DIO_DIRTY) {
+	if (dio->flags & IOMAP_DIO_USER_BACKED) {
 		bio_check_pages_dirty(bio);
 	} else {
 		bio_release_pages(bio, false);
@@ -333,7 +333,7 @@ static ssize_t iomap_dio_bio_iter_one(struct iomap_iter *iter,
 
 	if (dio->flags & IOMAP_DIO_WRITE)
 		task_io_account_write(ret);
-	else if (dio->flags & IOMAP_DIO_DIRTY)
+	else if (dio->flags & IOMAP_DIO_USER_BACKED)
 		bio_set_pages_dirty(bio);
 
 	/*
@@ -679,7 +679,7 @@ __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 			goto out_free_dio;
 
 		if (user_backed_iter(iter))
-			dio->flags |= IOMAP_DIO_DIRTY;
+			dio->flags |= IOMAP_DIO_USER_BACKED;
 
 		ret = kiocb_write_and_wait(iocb, iomi.len);
 		if (ret)
-- 
2.47.3

[PATCH 13/15] iomap: support ioends for direct reads

[Christoph Hellwig]

Support using the ioend structure to defer I/O completion for direct
reads in addition to writes.  This requires a check for the operation
to not merge reads and writes in iomap_ioend_can_merge.  This support
will be used for bounce buffered direct I/O reads that need to copy
data back to the user address space on read completion.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
---
 fs/iomap/ioend.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/fs/iomap/ioend.c b/fs/iomap/ioend.c
index 86f44922ed3b..800d12f45438 100644
--- a/fs/iomap/ioend.c
+++ b/fs/iomap/ioend.c
@@ -299,6 +299,14 @@ EXPORT_SYMBOL_GPL(iomap_finish_ioends);
 static bool iomap_ioend_can_merge(struct iomap_ioend *ioend,
 		struct iomap_ioend *next)
 {
+	/*
+	 * There is no point in merging reads as there is no completion
+	 * processing that can be easily batched up for them.
+	 */
+	if (bio_op(&ioend->io_bio) == REQ_OP_READ ||
+	    bio_op(&next->io_bio) == REQ_OP_READ)
+		return false;
+
 	if (ioend->io_bio.bi_status != next->io_bio.bi_status)
 		return false;
 	if (next->io_flags & IOMAP_IOEND_BOUNDARY)
-- 
2.47.3

[PATCH 14/15] iomap: add a flag to bounce buffer direct I/O

[Christoph Hellwig]

Add a new flag that request bounce buffering for direct I/O.  This is
needed to provide the stable pages requirement requested by devices
that need to calculate checksums or parity over the data and allows
file systems to properly work with things like T10 protection
information.  The implementation just calls out to the new bio bounce
buffering helpers to allocate a bounce buffer, which is used for
I/O and to copy to/from it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
---
 fs/iomap/direct-io.c  | 30 ++++++++++++++++++++----------
 include/linux/iomap.h |  9 +++++++++
 2 files changed, 29 insertions(+), 10 deletions(-)

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index eca7adda595a..9c572de0d596 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -215,7 +215,11 @@ static void __iomap_dio_bio_end_io(struct bio *bio, bool inline_completion)
 {
 	struct iomap_dio *dio = bio->bi_private;
 
-	if (dio->flags & IOMAP_DIO_USER_BACKED) {
+	if (dio->flags & IOMAP_DIO_BOUNCE) {
+		bio_iov_iter_unbounce(bio, !!dio->error,
+				dio->flags & IOMAP_DIO_USER_BACKED);
+		bio_put(bio);
+	} else if (dio->flags & IOMAP_DIO_USER_BACKED) {
 		bio_check_pages_dirty(bio);
 	} else {
 		bio_release_pages(bio, false);
@@ -303,12 +307,16 @@ static ssize_t iomap_dio_bio_iter_one(struct iomap_iter *iter,
 		struct iomap_dio *dio, loff_t pos, unsigned int alignment,
 		blk_opf_t op)
 {
+	unsigned int nr_vecs;
 	struct bio *bio;
 	ssize_t ret;
 
-	bio = iomap_dio_alloc_bio(iter, dio,
-			bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS),
-			op);
+	if (dio->flags & IOMAP_DIO_BOUNCE)
+		nr_vecs = bio_iov_bounce_nr_vecs(dio->submit.iter, op);
+	else
+		nr_vecs = bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS);
+
+	bio = iomap_dio_alloc_bio(iter, dio, nr_vecs, op);
 	fscrypt_set_bio_crypt_ctx(bio, iter->inode,
 			pos >> iter->inode->i_blkbits, GFP_KERNEL);
 	bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos);
@@ -317,7 +325,11 @@ static ssize_t iomap_dio_bio_iter_one(struct iomap_iter *iter,
 	bio->bi_private = dio;
 	bio->bi_end_io = iomap_dio_bio_end_io;
 
-	ret = bio_iov_iter_get_pages(bio, dio->submit.iter, alignment - 1);
+	if (dio->flags & IOMAP_DIO_BOUNCE)
+		ret = bio_iov_iter_bounce(bio, dio->submit.iter);
+	else
+		ret = bio_iov_iter_get_pages(bio, dio->submit.iter,
+					     alignment - 1);
 	if (unlikely(ret))
 		goto out_put_bio;
 	ret = bio->bi_iter.bi_size;
@@ -333,7 +345,8 @@ static ssize_t iomap_dio_bio_iter_one(struct iomap_iter *iter,
 
 	if (dio->flags & IOMAP_DIO_WRITE)
 		task_io_account_write(ret);
-	else if (dio->flags & IOMAP_DIO_USER_BACKED)
+	else if ((dio->flags & IOMAP_DIO_USER_BACKED) &&
+		 !(dio->flags & IOMAP_DIO_BOUNCE))
 		bio_set_pages_dirty(bio);
 
 	/*
@@ -662,7 +675,7 @@ __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 	dio->i_size = i_size_read(inode);
 	dio->dops = dops;
 	dio->error = 0;
-	dio->flags = 0;
+	dio->flags = dio_flags & (IOMAP_DIO_FSBLOCK_ALIGNED | IOMAP_DIO_BOUNCE);
 	dio->done_before = done_before;
 
 	dio->submit.iter = iter;
@@ -671,9 +684,6 @@ __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 	if (iocb->ki_flags & IOCB_NOWAIT)
 		iomi.flags |= IOMAP_NOWAIT;
 
-	if (dio_flags & IOMAP_DIO_FSBLOCK_ALIGNED)
-		dio->flags |= IOMAP_DIO_FSBLOCK_ALIGNED;
-
 	if (iov_iter_rw(iter) == READ) {
 		if (iomi.pos >= dio->i_size)
 			goto out_free_dio;
diff --git a/include/linux/iomap.h b/include/linux/iomap.h
index 520e967cb501..cf152f638665 100644
--- a/include/linux/iomap.h
+++ b/include/linux/iomap.h
@@ -562,6 +562,15 @@ struct iomap_dio_ops {
  */
 #define IOMAP_DIO_FSBLOCK_ALIGNED	(1 << 3)
 
+/*
+ * Bounce buffer instead of using zero copy access.
+ *
+ * This is needed if the device needs stable data to checksum or generate
+ * parity.  The file system must hook into the I/O submission and offload
+ * completions to user context for reads when this is set.
+ */
+#define IOMAP_DIO_BOUNCE		(1 << 4)
+
 ssize_t iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 		const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
 		unsigned int dio_flags, void *private, size_t done_before);
-- 
2.47.3

[PATCH 15/15] xfs: use bounce buffering direct I/O when the device requires stable pages

[Christoph Hellwig]

Fix direct I/O on devices that require stable pages by asking iomap
to bounce buffer.  To support this, ioends are used for direct reads
in this case to provide a user context for copying data back from the
bounce buffer.

This fixes qemu when used on devices using T10 protection information
and probably other cases like iSCSI using data digests.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
---
 fs/xfs/xfs_aops.c |  8 ++++++--
 fs/xfs/xfs_file.c | 41 ++++++++++++++++++++++++++++++++++++++---
 2 files changed, 44 insertions(+), 5 deletions(-)

diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c
index 56a544638491..c3c1e149fff4 100644
--- a/fs/xfs/xfs_aops.c
+++ b/fs/xfs/xfs_aops.c
@@ -103,7 +103,7 @@ xfs_ioend_put_open_zones(
  * IO write completion.
  */
 STATIC void
-xfs_end_ioend(
+xfs_end_ioend_write(
 	struct iomap_ioend	*ioend)
 {
 	struct xfs_inode	*ip = XFS_I(ioend->io_inode);
@@ -202,7 +202,11 @@ xfs_end_io(
 			io_list))) {
 		list_del_init(&ioend->io_list);
 		iomap_ioend_try_merge(ioend, &tmp);
-		xfs_end_ioend(ioend);
+		if (bio_op(&ioend->io_bio) == REQ_OP_READ)
+			iomap_finish_ioends(ioend,
+				blk_status_to_errno(ioend->io_bio.bi_status));
+		else
+			xfs_end_ioend_write(ioend);
 		cond_resched();
 	}
 }
diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index 7874cf745af3..f6cc63dcf961 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -224,12 +224,34 @@ xfs_ilock_iocb_for_write(
 	return 0;
 }
 
+/*
+ * Bounce buffering dio reads need a user context to copy back the data.
+ * Use an ioend to provide that.
+ */
+static void
+xfs_dio_read_bounce_submit_io(
+	const struct iomap_iter	*iter,
+	struct bio		*bio,
+	loff_t			file_offset)
+{
+	iomap_init_ioend(iter->inode, bio, file_offset, IOMAP_IOEND_DIRECT);
+	bio->bi_end_io = xfs_end_bio;
+	submit_bio(bio);
+}
+
+static const struct iomap_dio_ops xfs_dio_read_bounce_ops = {
+	.submit_io	= xfs_dio_read_bounce_submit_io,
+	.bio_set	= &iomap_ioend_bioset,
+};
+
 STATIC ssize_t
 xfs_file_dio_read(
 	struct kiocb		*iocb,
 	struct iov_iter		*to)
 {
 	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
+	unsigned int		dio_flags = 0;
+	const struct iomap_dio_ops *dio_ops = NULL;
 	ssize_t			ret;
 
 	trace_xfs_file_direct_read(iocb, to);
@@ -242,7 +264,12 @@ xfs_file_dio_read(
 	ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
 	if (ret)
 		return ret;
-	ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL, 0, NULL, 0);
+	if (mapping_stable_writes(iocb->ki_filp->f_mapping)) {
+		dio_ops = &xfs_dio_read_bounce_ops;
+		dio_flags |= IOMAP_DIO_BOUNCE;
+	}
+	ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, dio_ops, dio_flags,
+			NULL, 0);
 	xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 
 	return ret;
@@ -703,6 +730,8 @@ xfs_file_dio_write_aligned(
 		xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
 		iolock = XFS_IOLOCK_SHARED;
 	}
+	if (mapping_stable_writes(iocb->ki_filp->f_mapping))
+		dio_flags |= IOMAP_DIO_BOUNCE;
 	trace_xfs_file_direct_write(iocb, from);
 	ret = iomap_dio_rw(iocb, from, ops, dops, dio_flags, ac, 0);
 out_unlock:
@@ -750,6 +779,7 @@ xfs_file_dio_write_atomic(
 {
 	unsigned int		iolock = XFS_IOLOCK_SHARED;
 	ssize_t			ret, ocount = iov_iter_count(from);
+	unsigned int		dio_flags = 0;
 	const struct iomap_ops	*dops;
 
 	/*
@@ -777,8 +807,10 @@ xfs_file_dio_write_atomic(
 	}
 
 	trace_xfs_file_direct_write(iocb, from);
-	ret = iomap_dio_rw(iocb, from, dops, &xfs_dio_write_ops,
-			0, NULL, 0);
+	if (mapping_stable_writes(iocb->ki_filp->f_mapping))
+		dio_flags |= IOMAP_DIO_BOUNCE;
+	ret = iomap_dio_rw(iocb, from, dops, &xfs_dio_write_ops, dio_flags,
+			NULL, 0);
 
 	/*
 	 * The retry mechanism is based on the ->iomap_begin method returning
@@ -867,6 +899,9 @@ xfs_file_dio_write_unaligned(
 	if (flags & IOMAP_DIO_FORCE_WAIT)
 		inode_dio_wait(VFS_I(ip));
 
+	if (mapping_stable_writes(iocb->ki_filp->f_mapping))
+		flags |= IOMAP_DIO_BOUNCE;
+
 	trace_xfs_file_direct_write(iocb, from);
 	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
 			   &xfs_dio_write_ops, flags, NULL, 0);
-- 
2.47.3

Re: bounce buffer direct I/O when stable pages are required v3

[Anuj gupta]

As Keith suggested, here are the QD1 latency numbers (in usec)

Intel Optane:

Sequential write
  | size | zero copy  |  bounce   |
  +------+------------+-----------+
  |   4k |    8.91    |    9.09   |
  |  64K |    35.98   |    39.9   |
  |   1M |    341.96  |    531.51 |
  +------+-------------------------+

Sequential read
  | size | zero copy  |  bounce   |
  +------+------------+-----------+
  |   4k |    7.18    |    14.28  |
  |  64K |    36.4    |    95.61  |
  |   1M |    206.38  |    258.66 |
  +------+-------------------------+

Re: bounce buffer direct I/O when stable pages are required v3

[Christoph Hellwig]

On Mon, Jan 26, 2026 at 04:24:03PM +0530, Anuj gupta wrote:
> As Keith suggested, here are the QD1 latency numbers (in usec)

Thanks a gain for the benchmarks!

I'd be curious what improvement you see with the iomap-pi series on
Optane, as that drops one of the context switches on read again,
and the less efficiently managed one at that:

git://git.infradead.org/users/hch/misc.git iomap-pi
https://git.infradead.org/?p=users/hch/misc.git;a=shortlog;h=refs/heads/iomap-pi

Otherwise the only thing we can do to get data integrity and performance
is better interfaces.  I think for reads we could do that relatively
easily with a version of Joanne's kernel-managed buffer rings that can
only be mapped into userspace read-only.  Writes will be more difficult
for anything that isn't a trust-worthy kernel provided buffer
unfortunately, but then again the write degradation is less.

Re: bounce buffer direct I/O when stable pages are required v3

[Carlos Maiolino]

On Mon, Jan 26, 2026 at 06:53:31AM +0100, Christoph Hellwig wrote:
> Hi all,
> 
> [note to maintainers:  we're ready to merge I think, and Christian
> already said he'd do on Friday.  If acceptable to everyone I'd like
> to merge it through the block tree, or topic branch in it due to
> pending work on top of this]

FWIW, no problems from my side with this approach.

Carlos

> 
> this series tries to address the problem that under I/O pages can be
> modified during direct I/O, even when the device or file system require
> stable pages during I/O to calculate checksums, parity or data
> operations.  It does so by adding block layer helpers to bounce buffer
> an iov_iter into a bio, then wires that up in iomap and ultimately
> XFS.
> 
> The reason that the file system even needs to know about it, is because
> reads need a user context to copy the data back, and the infrastructure
> to defer ioends to a workqueue currently sits in XFS.  I'm going to look
> into moving that into ioend and enabling it for other file systems.
> Additionally btrfs already has it's own infrastructure for this, and
> actually an urgent need to bounce buffer, so this should be useful there
> and could be wire up easily.  In fact the idea comes from patches by
> Qu that did this in btrfs.
> 
> This patch fixes all but one xfstests failures on T10 PI capable devices
> (generic/095 seems to have issues with a mix of mmap and splice still,
> I'm looking into that separate), and make qemu VMs running Windows,
> or Linux with swap enabled fine on an XFS file on a device using PI.
> 
> Performance numbers on my (not exactly state of the art) NVMe PI test
> setup:
> 
>   Sequential reads using io_uring, QD=16.
>   Bandwidth and CPU usage (usr/sys):
> 
>   | size |        zero copy         |          bounce          |
>   +------+--------------------------+--------------------------+
>   |   4k | 1316MiB/s (12.65/55.40%) | 1081MiB/s (11.76/49.78%) |
>   |  64K | 3370MiB/s ( 5.46/18.20%) | 3365MiB/s ( 4.47/15.68%) |
>   |   1M | 3401MiB/s ( 0.76/23.05%) | 3400MiB/s ( 0.80/09.06%) |
>   +------+--------------------------+--------------------------+
> 
>   Sequential writes using io_uring, QD=16.
>   Bandwidth and CPU usage (usr/sys):
> 
>   | size |        zero copy         |          bounce          |
>   +------+--------------------------+--------------------------+
>   |   4k |  882MiB/s (11.83/33.88%) |  750MiB/s (10.53/34.08%) |
>   |  64K | 2009MiB/s ( 7.33/15.80%) | 2007MiB/s ( 7.47/24.71%) |
>   |   1M | 1992MiB/s ( 7.26/ 9.13%) | 1992MiB/s ( 9.21/19.11%) |
>   +------+--------------------------+--------------------------+
> 
> Note that the 64k read numbers look really odd to me for the baseline
> zero copy case, but are reproducible over many repeated runs.
> 
> The bounce read numbers should further improve when moving the PI
> validation to the file system and removing the double context switch,
> which I have patches for that will sent out soon.
> 
> Changes since v2:
>  - add a BIO_MAX_SIZE constant and use it
>  - remove a pointless repeated page_folio call
>  - fix a comment typo
>  - add a new comment about copying to a pinned iter
> 
> Changes since v1:
>  - spelling fixes
>  - add more details to some commit messages
>  - add a new code comment about freeing the bio early in the I/O
>    completion handler
> 
> Diffstat:
>  block/bio.c               |  332 ++++++++++++++++++++++++++++------------------
>  block/blk-lib.c           |    9 -
>  block/blk-merge.c         |    8 -
>  block/blk.h               |   11 -
>  fs/iomap/direct-io.c      |  191 ++++++++++++++------------
>  fs/iomap/ioend.c          |    8 +
>  fs/xfs/xfs_aops.c         |    8 -
>  fs/xfs/xfs_file.c         |   41 +++++
>  include/linux/bio.h       |   26 +++
>  include/linux/blk_types.h |    3 
>  include/linux/iomap.h     |    9 +
>  include/linux/uio.h       |    3 
>  lib/iov_iter.c            |   98 +++++++++++++
>  13 files changed, 507 insertions(+), 240 deletions(-)
> 

Re: bounce buffer direct I/O when stable pages are required v3

[Jens Axboe]

On Mon, 26 Jan 2026 06:53:31 +0100, Christoph Hellwig wrote:
> [note to maintainers:  we're ready to merge I think, and Christian
> already said he'd do on Friday.  If acceptable to everyone I'd like
> to merge it through the block tree, or topic branch in it due to
> pending work on top of this]
> 
> this series tries to address the problem that under I/O pages can be
> modified during direct I/O, even when the device or file system require
> stable pages during I/O to calculate checksums, parity or data
> operations.  It does so by adding block layer helpers to bounce buffer
> an iov_iter into a bio, then wires that up in iomap and ultimately
> XFS.
> 
> [...]

Applied, thanks!

[01/15] block: add a BIO_MAX_SIZE constant and use it
        commit: fa0bdd45d7e3703826ea75f5fe3359865d75c319
[02/15] block: refactor get_contig_folio_len
        commit: 4d77007d42fd4f44c2f5a1555603df53e16a1362
[03/15] block: open code bio_add_page and fix handling of mismatching P2P ranges
        commit: 12da89e8844ae16e86b75a32b34a4f0b0525f453
[04/15] iov_iter: extract a iov_iter_extract_bvecs helper from bio code
        commit: 91b73c458182801a8c9cf6135335e064567d1013
[05/15] block: remove bio_release_page
        commit: 301f5356521ed90f72a67797156d75093aac786f
[06/15] block: add helpers to bounce buffer an iov_iter into bios
        commit: 8dd5e7c75d7bb2635c7efd219ff20693fc24096a
[07/15] iomap: fix submission side handling of completion side errors
        commit: 4ad357e39b2ecd5da7bcc7e840ee24d179593cd5
[08/15] iomap: simplify iomap_dio_bio_iter
        commit: 6e7a6c80198ead08b11aa6cdc92e60a42fc5895f
[09/15] iomap: split out the per-bio logic from iomap_dio_bio_iter
        commit: 2631c94602297090febd8f93d6f96d9d2045466d
[10/15] iomap: share code between iomap_dio_bio_end_io and iomap_finish_ioend_direct
        commit: e2fcff5bb4c48bf602082e5a1428ff7328f7558f
[11/15] iomap: free the bio before completing the dio
        commit: 45cec0de6c8973660da279e44b24d37af49daeb6
[12/15] iomap: rename IOMAP_DIO_DIRTY to IOMAP_DIO_USER_BACKED
        commit: c96b8b220271024c04289d6d9779dc2ccbd12be2
[13/15] iomap: support ioends for direct reads
        commit: d969bd72cf6835a4c915b326feb92c7597a46d98
[14/15] iomap: add a flag to bounce buffer direct I/O
        commit: c9d114846b380fec1093b7bca91ee5a8cd7b575d
[15/15] xfs: use bounce buffering direct I/O when the device requires stable pages
        commit: 3373503df025ab6c9a8ad2ce6b7febd2eb3c99dc

Best regards,
-- 
Jens Axboe

Re: bounce buffer direct I/O when stable pages are required v3

[Jens Axboe]

On 1/25/26 10:53 PM, Christoph Hellwig wrote:
> Hi all,
> 
> [note to maintainers:  we're ready to merge I think, and Christian
> already said he'd do on Friday.  If acceptable to everyone I'd like
> to merge it through the block tree, or topic branch in it due to
> pending work on top of this]

Queued up in a fork off for-7.0/block, for-7.0/block-stable-pages.

-- 
Jens Axboe