[PATCH v6 0/2] io_uring/rsrc: coalescing multi-hugepage registered buffers

[PATCH v6 0/2] io_uring/rsrc: coalescing multi-hugepage registered buffers

[Chenliang Li]

Registered buffers are stored and processed in the form of bvec array,
each bvec element typically points to a PAGE_SIZE page but can also work
with hugepages. Specifically, a buffer consisting of a hugepage is
coalesced to use only one hugepage bvec entry during registration.
This coalescing feature helps to save both the space and DMA-mapping time.

However, currently the coalescing feature doesn't work for multi-hugepage
buffers. For a buffer with several 2M hugepages, we still split it into
thousands of 4K page bvec entries while in fact, we can just use a
handful of hugepage bvecs.

This patch series enables coalescing registered buffers with more than
one hugepages. It optimizes the DMA-mapping time and saves memory for 
these kind of buffers.

Testing:

The hugepage fixed buffer I/O can be tested using fio without
modification. The fio command used in the following test is given
in [1]. There's also a liburing testcase in [2]. Also, the system
should have enough hugepages available before testing.

Perf diff of 8M(4 * 2M hugepages) fio randread test:

Before          After           Symbol
.....................................................
5.88%                           [k] __blk_rq_map_sg
3.98%           -3.95%          [k] dma_direct_map_sg
2.47%                           [k] dma_pool_alloc
1.37%           -1.36%          [k] sg_next
                +0.28%          [k] dma_map_page_attrs

Perf diff of 8M fio randwrite test:

Before          After           Symbol
......................................................
2.80%                           [k] __blk_rq_map_sg
1.74%                           [k] dma_direct_map_sg
1.61%                           [k] dma_pool_alloc
0.67%                           [k] sg_next
                +0.04%          [k] dma_map_page_attrs

The first patch prepares for adding the multi-hugepage coalescing
by storing folio_shift and folio_mask into imu, the 2nd patch
enables the feature.

---

Changes since v5:
- Reshuffle the patchset to avoid unused funtion warnings.
- Store head page of the folio and use folio offset to get rid of
  branching on bvec setups.
- Add restrictions for non border-aligned folios.
- Remove unnecessary folio_size field in io_imu_folio_data struct.

v5 : https://lore.kernel.org/io-uring/20240708021426.2217-1-cliang01.li@samsung.com/T/#t

[1]
fio -iodepth=64 -rw=randread -direct=1 -ioengine=io_uring \
-bs=8M -numjobs=1 -group_reporting -mem=shmhuge -fixedbufs -hugepage-size=2M \
-filename=/dev/nvme0n1 -runtime=10s -name=test1

[2]
https://lore.kernel.org/io-uring/20240514051343.582556-1-cliang01.li@samsung.com/T/#u

Chenliang Li (2):
  io_uring/rsrc: store folio shift and mask into imu
  io_uring/rsrc: enable multi-hugepage buffer coalescing

 io_uring/rsrc.c | 154 +++++++++++++++++++++++++++++++++++-------------
 io_uring/rsrc.h |  10 ++++
 2 files changed, 123 insertions(+), 41 deletions(-)


base-commit: ad00e629145b2b9f0d78aa46e204a9df7d628978
-- 
2.34.1

[PATCH v6 1/2] io_uring/rsrc: store folio shift and mask into imu

[Chenliang Li]

Store the folio shift and folio mask into imu struct and use it in
iov_iter adjust, as we will have non PAGE_SIZE'd chunks if a
multi-hugepage buffer get coalesced.

Signed-off-by: Chenliang Li <cliang01.li@samsung.com>
Reviewed-by: Anuj Gupta <anuj20.g@samsung.com>
---
 io_uring/rsrc.c | 15 ++++++---------
 io_uring/rsrc.h |  2 ++
 2 files changed, 8 insertions(+), 9 deletions(-)

diff --git a/io_uring/rsrc.c b/io_uring/rsrc.c
index 60c00144471a..0d6cda92ba46 100644
--- a/io_uring/rsrc.c
+++ b/io_uring/rsrc.c
@@ -915,6 +915,8 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
 	imu->ubuf = (unsigned long) iov->iov_base;
 	imu->ubuf_end = imu->ubuf + iov->iov_len;
 	imu->nr_bvecs = nr_pages;
+	imu->folio_shift = PAGE_SHIFT;
+	imu->folio_mask = PAGE_MASK;
 	*pimu = imu;
 	ret = 0;
 
@@ -1031,23 +1033,18 @@ int io_import_fixed(int ddir, struct iov_iter *iter,
 		 * we know that:
 		 *
 		 * 1) it's a BVEC iter, we set it up
-		 * 2) all bvecs are PAGE_SIZE in size, except potentially the
+		 * 2) all bvecs are the same in size, except potentially the
 		 *    first and last bvec
 		 *
 		 * So just find our index, and adjust the iterator afterwards.
 		 * If the offset is within the first bvec (or the whole first
 		 * bvec, just use iov_iter_advance(). This makes it easier
 		 * since we can just skip the first segment, which may not
-		 * be PAGE_SIZE aligned.
+		 * be folio_size aligned.
 		 */
 		const struct bio_vec *bvec = imu->bvec;
 
 		if (offset < bvec->bv_len) {
-			/*
-			 * Note, huge pages buffers consists of one large
-			 * bvec entry and should always go this way. The other
-			 * branch doesn't expect non PAGE_SIZE'd chunks.
-			 */
 			iter->bvec = bvec;
 			iter->nr_segs = bvec->bv_len;
 			iter->count -= offset;
@@ -1057,12 +1054,12 @@ int io_import_fixed(int ddir, struct iov_iter *iter,
 
 			/* skip first vec */
 			offset -= bvec->bv_len;
-			seg_skip = 1 + (offset >> PAGE_SHIFT);
+			seg_skip = 1 + (offset >> imu->folio_shift);
 
 			iter->bvec = bvec + seg_skip;
 			iter->nr_segs -= seg_skip;
 			iter->count -= bvec->bv_len + offset;
-			iter->iov_offset = offset & ~PAGE_MASK;
+			iter->iov_offset = offset & ~imu->folio_mask;
 		}
 	}
 
diff --git a/io_uring/rsrc.h b/io_uring/rsrc.h
index c032ca3436ca..8b029c53d325 100644
--- a/io_uring/rsrc.h
+++ b/io_uring/rsrc.h
@@ -46,7 +46,9 @@ struct io_mapped_ubuf {
 	u64		ubuf;
 	u64		ubuf_end;
 	unsigned int	nr_bvecs;
+	unsigned int	folio_shift;
 	unsigned long	acct_pages;
+	unsigned long	folio_mask;
 	struct bio_vec	bvec[] __counted_by(nr_bvecs);
 };
 
-- 
2.34.1

[PATCH v6 2/2] io_uring/rsrc: enable multi-hugepage buffer coalescing

[Chenliang Li]

Add support for checking and coalescing multi-hugepage-backed fixed
buffers. The coalescing optimizes both time and space consumption caused
by mapping and storing multi-hugepage fixed buffers.

A coalescable multi-hugepage buffer should fully cover its folios
(except potentially the first and last one), and these folios should
have the same size. These requirements are for easier processing later,
also we need same size'd chunks in io_import_fixed for fast iov_iter
adjust.

Signed-off-by: Chenliang Li <cliang01.li@samsung.com>
---
 io_uring/rsrc.c | 139 +++++++++++++++++++++++++++++++++++++-----------
 io_uring/rsrc.h |   8 +++
 2 files changed, 115 insertions(+), 32 deletions(-)

diff --git a/io_uring/rsrc.c b/io_uring/rsrc.c
index 0d6cda92ba46..bba7d2913aa9 100644
--- a/io_uring/rsrc.c
+++ b/io_uring/rsrc.c
@@ -849,6 +849,103 @@ static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
 	return ret;
 }
 
+static bool io_do_coalesce_buffer(struct page ***pages, int *nr_pages,
+				struct io_imu_folio_data *data, int nr_folios)
+{
+	struct page **page_array = *pages, **new_array = NULL;
+	int nr_pages_left = *nr_pages, i, j;
+
+	/* Store head pages only*/
+	new_array = kvmalloc_array(nr_folios, sizeof(struct page *),
+					GFP_KERNEL);
+	if (!new_array)
+		return false;
+
+	new_array[0] = compound_head(page_array[0]);
+	/*
+	 * The pages are bound to the folio, it doesn't
+	 * actually unpin them but drops all but one reference,
+	 * which is usually put down by io_buffer_unmap().
+	 * Note, needs a better helper.
+	 */
+	if (data->nr_pages_head > 1)
+		unpin_user_pages(&page_array[1], data->nr_pages_head - 1);
+
+	j = data->nr_pages_head;
+	nr_pages_left -= data->nr_pages_head;
+	for (i = 1; i < nr_folios; i++) {
+		unsigned int nr_unpin;
+
+		new_array[i] = page_array[j];
+		nr_unpin = min_t(unsigned int, nr_pages_left - 1,
+					data->nr_pages_mid - 1);
+		if (nr_unpin)
+			unpin_user_pages(&page_array[j+1], nr_unpin);
+		j += data->nr_pages_mid;
+		nr_pages_left -= data->nr_pages_mid;
+	}
+	kvfree(page_array);
+	*pages = new_array;
+	*nr_pages = nr_folios;
+	return true;
+}
+
+static bool io_try_coalesce_buffer(struct page ***pages, int *nr_pages,
+					 struct io_imu_folio_data *data)
+{
+	struct page **page_array = *pages;
+	struct folio *folio = page_folio(page_array[0]);
+	unsigned int count = 1, nr_folios = 1;
+	int i;
+
+	if (*nr_pages <= 1)
+		return false;
+
+	data->nr_pages_mid = folio_nr_pages(folio);
+	if (data->nr_pages_mid == 1)
+		return false;
+
+	data->folio_shift = folio_shift(folio);
+	/*
+	 * Check if pages are contiguous inside a folio, and all folios have
+	 * the same page count except for the head and tail.
+	 */
+	for (i = 1; i < *nr_pages; i++) {
+		if (page_folio(page_array[i]) == folio &&
+			page_array[i] == page_array[i-1] + 1) {
+			count++;
+			continue;
+		}
+
+		if (nr_folios == 1) {
+			if (folio_page_idx(folio, page_array[i-1])
+				!= data->nr_pages_mid - 1)
+				return false;
+
+			data->nr_pages_head = count;
+		} else if (count != data->nr_pages_mid) {
+			return false;
+		}
+
+		folio = page_folio(page_array[i]);
+		if (folio_size(folio) != (1UL << data->folio_shift) ||
+			folio_page_idx(folio, page_array[i]) != 0)
+			return false;
+
+		count = 1;
+		nr_folios++;
+	}
+	if (nr_folios == 1) {
+		if (folio_page_idx(folio, page_array[i-1])
+			!= data->nr_pages_mid - 1)
+			return false;
+
+		data->nr_pages_head = count;
+	}
+
+	return io_do_coalesce_buffer(pages, nr_pages, data, nr_folios);
+}
+
 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
 				  struct io_mapped_ubuf **pimu,
 				  struct page **last_hpage)
@@ -858,7 +955,8 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
 	unsigned long off;
 	size_t size;
 	int ret, nr_pages, i;
-	struct folio *folio = NULL;
+	struct io_imu_folio_data data;
+	bool coalesced;
 
 	*pimu = (struct io_mapped_ubuf *)&dummy_ubuf;
 	if (!iov->iov_base)
@@ -873,31 +971,8 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
 		goto done;
 	}
 
-	/* If it's a huge page, try to coalesce them into a single bvec entry */
-	if (nr_pages > 1) {
-		folio = page_folio(pages[0]);
-		for (i = 1; i < nr_pages; i++) {
-			/*
-			 * Pages must be consecutive and on the same folio for
-			 * this to work
-			 */
-			if (page_folio(pages[i]) != folio ||
-			    pages[i] != pages[i - 1] + 1) {
-				folio = NULL;
-				break;
-			}
-		}
-		if (folio) {
-			/*
-			 * The pages are bound to the folio, it doesn't
-			 * actually unpin them but drops all but one reference,
-			 * which is usually put down by io_buffer_unmap().
-			 * Note, needs a better helper.
-			 */
-			unpin_user_pages(&pages[1], nr_pages - 1);
-			nr_pages = 1;
-		}
-	}
+	/* If it's huge page(s), try to coalesce them into fewer bvec entries */
+	coalesced = io_try_coalesce_buffer(&pages, &nr_pages, &data);
 
 	imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
 	if (!imu)
@@ -909,7 +984,6 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
 		goto done;
 	}
 
-	off = (unsigned long) iov->iov_base & ~PAGE_MASK;
 	size = iov->iov_len;
 	/* store original address for later verification */
 	imu->ubuf = (unsigned long) iov->iov_base;
@@ -917,17 +991,18 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
 	imu->nr_bvecs = nr_pages;
 	imu->folio_shift = PAGE_SHIFT;
 	imu->folio_mask = PAGE_MASK;
+	if (coalesced) {
+		imu->folio_shift = data.folio_shift;
+		imu->folio_mask = ~((1UL << data.folio_shift) - 1);
+	}
+	off = (unsigned long) iov->iov_base & ~imu->folio_mask;
 	*pimu = imu;
 	ret = 0;
 
-	if (folio) {
-		bvec_set_page(&imu->bvec[0], pages[0], size, off);
-		goto done;
-	}
 	for (i = 0; i < nr_pages; i++) {
 		size_t vec_len;
 
-		vec_len = min_t(size_t, size, PAGE_SIZE - off);
+		vec_len = min_t(size_t, size, (1UL << imu->folio_shift) - off);
 		bvec_set_page(&imu->bvec[i], pages[i], vec_len, off);
 		off = 0;
 		size -= vec_len;
diff --git a/io_uring/rsrc.h b/io_uring/rsrc.h
index 8b029c53d325..048fd6959f7a 100644
--- a/io_uring/rsrc.h
+++ b/io_uring/rsrc.h
@@ -52,6 +52,14 @@ struct io_mapped_ubuf {
 	struct bio_vec	bvec[] __counted_by(nr_bvecs);
 };
 
+struct io_imu_folio_data {
+	/* Head folio can be partially included in the fixed buf */
+	unsigned int	nr_pages_head;
+	/* For non-head/tail folios, has to be fully included */
+	unsigned int	nr_pages_mid;
+	unsigned int	folio_shift;
+};
+
 void io_rsrc_node_ref_zero(struct io_rsrc_node *node);
 void io_rsrc_node_destroy(struct io_ring_ctx *ctx, struct io_rsrc_node *ref_node);
 struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
-- 
2.34.1

Re: [PATCH v6 2/2] io_uring/rsrc: enable multi-hugepage buffer coalescing

[Pavel Begunkov]

On 7/16/24 07:08, Chenliang Li wrote:
> Add support for checking and coalescing multi-hugepage-backed fixed
> buffers. The coalescing optimizes both time and space consumption caused
> by mapping and storing multi-hugepage fixed buffers.
> 
> A coalescable multi-hugepage buffer should fully cover its folios
> (except potentially the first and last one), and these folios should
> have the same size. These requirements are for easier processing later,
> also we need same size'd chunks in io_import_fixed for fast iov_iter
> adjust.
> 
> Signed-off-by: Chenliang Li <cliang01.li@samsung.com>
> ---

The series looks good apart from the comment below, I'll run some
tests this week, I think you sent something for liburing/test

> +static bool io_try_coalesce_buffer(struct page ***pages, int *nr_pages,
> +					 struct io_imu_folio_data *data)
> +{
> +	struct page **page_array = *pages;
> +	struct folio *folio = page_folio(page_array[0]);
> +	unsigned int count = 1, nr_folios = 1;
> +	int i;
> +
> +	if (*nr_pages <= 1)
> +		return false;
> +
> +	data->nr_pages_mid = folio_nr_pages(folio);
> +	if (data->nr_pages_mid == 1)
> +		return false;
> +
> +	data->folio_shift = folio_shift(folio);
> +	/*
> +	 * Check if pages are contiguous inside a folio, and all folios have
> +	 * the same page count except for the head and tail.
> +	 */
> +	for (i = 1; i < *nr_pages; i++) {
> +		if (page_folio(page_array[i]) == folio &&
> +			page_array[i] == page_array[i-1] + 1) {
> +			count++;
> +			continue;
> +		}
> +
> +		if (nr_folios == 1) {
> +			if (folio_page_idx(folio, page_array[i-1])
> +				!= data->nr_pages_mid - 1)

Code style, comparison (i.e. "!=") should be on the previous line

> +				return false;
> +
> +			data->nr_pages_head = count;
> +		} else if (count != data->nr_pages_mid) {
> +			return false;
> +		}
> +
> +		folio = page_folio(page_array[i]);
> +		if (folio_size(folio) != (1UL << data->folio_shift) ||
> +			folio_page_idx(folio, page_array[i]) != 0)
> +			return false;
> +
> +		count = 1;
> +		nr_folios++;
> +	}
> +	if (nr_folios == 1) {
> +		if (folio_page_idx(folio, page_array[i-1])
> +			!= data->nr_pages_mid - 1)

That's too restrictive and would be a regression. We currently allow
registrations fully falling under a single huge page but not matching
the right side. E.g. for a [N; M) huge page, you're allowed to register
[N + 4KB, M - 4KB). We should just remove this check.

And same comment about code style.


> +			return false;
> +
> +		data->nr_pages_head = count;
> +	}
> +
> +	return io_do_coalesce_buffer(pages, nr_pages, data, nr_folios);
> +}
> +
-- 
Pavel Begunkov