[PATCH V9 00/19] virtio_ring in order support

[PATCH V9 00/19] virtio_ring in order support

[Jason Wang]

Hello all:

This sereis tries to implement the VIRTIO_F_IN_ORDER to
virtio_ring. This is done by introducing virtqueue ops so we can
implement separate helpers for different virtqueue layout/features
then the in-order were implemented on top.

Tests shows 2%-19% imporvment with packed virtqueue PPS with KVM guest
vhost-net/testpmd on the host.

Changes since V8:

- Rebase to vhost.git linux-next branch
- Fix issue when indirect descriptor is disabled
- use UINT_MAX instead of vq->num to indicate no batching is pending
- Tweak the comments for used_entry
- Various coding style tweaks and typo fixes
- Try to use in_order varianet in in_order ops

Changes since v7:

- Rebase on vhost.git linux-next branch
- Tweak the comment to explain the usage of free_head

Changes since V6:

- Rebase on vhost.git linux-next branch
- Fix poking packed virtqueue in more_used_split_in_order()
- Fix calling detach_buf_packed_in_order() unconditonally in
  virtqueue_detach_unused_buf_packed()
- Typo and indentation fixes
- Fix wrong changelog of patch 7

Changes since V5:

- rebase on vhost.git linux-next branch
- reorder the total_len to reduce memory comsuming

Changes since V4:

- Fix build error when DEBUG is enabled
- Fix function duplications
- Remove unnecessary new lines

Changes since V3:

- Re-benchmark with the recent vhost-net in order support
- Rename the batched used id and length
- Other minor tweaks

Changes since V2:

- Fix build warning when DEBUG is enabled

Changes since V1:

- use const global array of function pointers to avoid indirect
  branches to eliminate retpoline when mitigation is enabled
- fix used length calculation when processing used ids in a batch
- fix sparse warnings

Jason Wang (19):
  virtio_ring: rename virtqueue_reinit_xxx to virtqueue_reset_xxx()
  virtio_ring: switch to use vring_virtqueue in virtqueue_poll variants
  virtio_ring: unify logic of virtqueue_poll() and more_used()
  virtio_ring: switch to use vring_virtqueue for virtqueue resize
    variants
  virtio_ring: switch to use vring_virtqueue for virtqueue_kick_prepare
    variants
  virtio_ring: switch to use vring_virtqueue for virtqueue_add variants
  virtio: switch to use vring_virtqueue for virtqueue_get variants
  virtio_ring: switch to use vring_virtqueue for enable_cb_prepare
    variants
  virtio_ring: use vring_virtqueue for enable_cb_delayed variants
  virtio_ring: switch to use vring_virtqueue for disable_cb variants
  virtio_ring: switch to use vring_virtqueue for detach_unused_buf
    variants
  virtio_ring: switch to use unsigned int for virtqueue_poll_packed()
  virtio_ring: introduce virtqueue ops
  virtio_ring: determine descriptor flags at one time
  virtio_ring: factor out core logic of buffer detaching
  virtio_ring: factor out core logic for updating last_used_idx
  virtio_ring: factor out split indirect detaching logic
  virtio_ring: factor out split detaching logic
  virtio_ring: add in order support

 drivers/virtio/virtio_ring.c | 930 +++++++++++++++++++++++++++--------
 1 file changed, 716 insertions(+), 214 deletions(-)

-- 
2.31.1

[PATCH V9 01/19] virtio_ring: rename virtqueue_reinit_xxx to virtqueue_reset_xxx()

[Jason Wang]

To be consistent with virtqueue_reset().

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index ddab68959671..cc1b8bbc6774 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -1009,7 +1009,7 @@ static void virtqueue_vring_init_split(struct vring_virtqueue_split *vring_split
 	}
 }
 
-static void virtqueue_reinit_split(struct vring_virtqueue *vq)
+static void virtqueue_reset_split(struct vring_virtqueue *vq)
 {
 	int num;
 
@@ -1253,7 +1253,7 @@ static int virtqueue_resize_split(struct virtqueue *_vq, u32 num)
 err_state_extra:
 	vring_free_split(&vring_split, vdev, vq->map);
 err:
-	virtqueue_reinit_split(vq);
+	virtqueue_reset_split(vq);
 	return -ENOMEM;
 }
 
@@ -2091,7 +2091,7 @@ static void virtqueue_vring_attach_packed(struct vring_virtqueue *vq,
 	vq->free_head = 0;
 }
 
-static void virtqueue_reinit_packed(struct vring_virtqueue *vq)
+static void virtqueue_reset_packed(struct vring_virtqueue *vq)
 {
 	memset(vq->packed.vring.device, 0, vq->packed.event_size_in_bytes);
 	memset(vq->packed.vring.driver, 0, vq->packed.event_size_in_bytes);
@@ -2218,7 +2218,7 @@ static int virtqueue_resize_packed(struct virtqueue *_vq, u32 num)
 err_state_extra:
 	vring_free_packed(&vring_packed, vdev, vq->map);
 err_ring:
-	virtqueue_reinit_packed(vq);
+	virtqueue_reset_packed(vq);
 	return -ENOMEM;
 }
 
@@ -2860,9 +2860,9 @@ int virtqueue_reset(struct virtqueue *_vq,
 		recycle_done(_vq);
 
 	if (vq->packed_ring)
-		virtqueue_reinit_packed(vq);
+		virtqueue_reset_packed(vq);
 	else
-		virtqueue_reinit_split(vq);
+		virtqueue_reset_split(vq);
 
 	return virtqueue_enable_after_reset(_vq);
 }
-- 
2.31.1

[PATCH V9 02/19] virtio_ring: switch to use vring_virtqueue in virtqueue_poll variants

[Jason Wang]

Those variants are used internally so let's switch to use
vring_virtqueue as parameter to be consistent with other internal
virtqueue helpers.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 14 ++++++--------
 1 file changed, 6 insertions(+), 8 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index cc1b8bbc6774..d5a317cb885f 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -919,11 +919,10 @@ static unsigned int virtqueue_enable_cb_prepare_split(struct virtqueue *_vq)
 	return last_used_idx;
 }
 
-static bool virtqueue_poll_split(struct virtqueue *_vq, unsigned int last_used_idx)
+static bool virtqueue_poll_split(struct vring_virtqueue *vq,
+				 unsigned int last_used_idx)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
-
-	return (u16)last_used_idx != virtio16_to_cpu(_vq->vdev,
+	return (u16)last_used_idx != virtio16_to_cpu(vq->vq.vdev,
 			vq->split.vring.used->idx);
 }
 
@@ -1844,9 +1843,8 @@ static unsigned int virtqueue_enable_cb_prepare_packed(struct virtqueue *_vq)
 	return vq->last_used_idx;
 }
 
-static bool virtqueue_poll_packed(struct virtqueue *_vq, u16 off_wrap)
+static bool virtqueue_poll_packed(struct vring_virtqueue *vq, u16 off_wrap)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
 	bool wrap_counter;
 	u16 used_idx;
 
@@ -2611,8 +2609,8 @@ bool virtqueue_poll(struct virtqueue *_vq, unsigned int last_used_idx)
 		return false;
 
 	virtio_mb(vq->weak_barriers);
-	return vq->packed_ring ? virtqueue_poll_packed(_vq, last_used_idx) :
-				 virtqueue_poll_split(_vq, last_used_idx);
+	return vq->packed_ring ? virtqueue_poll_packed(vq, last_used_idx) :
+				 virtqueue_poll_split(vq, last_used_idx);
 }
 EXPORT_SYMBOL_GPL(virtqueue_poll);
 
-- 
2.31.1

[PATCH V9 03/19] virtio_ring: unify logic of virtqueue_poll() and more_used()

[Jason Wang]

This patch unifies the logic of virtqueue_poll() and more_used() for
better code reusing and ease the future in order implementation.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 48 +++++++++++++++---------------------
 1 file changed, 20 insertions(+), 28 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index d5a317cb885f..a32050e1aa47 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -806,12 +806,18 @@ static void detach_buf_split(struct vring_virtqueue *vq, unsigned int head,
 	}
 }
 
-static bool more_used_split(const struct vring_virtqueue *vq)
+static bool virtqueue_poll_split(const struct vring_virtqueue *vq,
+				 unsigned int last_used_idx)
 {
-	return vq->last_used_idx != virtio16_to_cpu(vq->vq.vdev,
+	return (u16)last_used_idx != virtio16_to_cpu(vq->vq.vdev,
 			vq->split.vring.used->idx);
 }
 
+static bool more_used_split(const struct vring_virtqueue *vq)
+{
+	return virtqueue_poll_split(vq, vq->last_used_idx);
+}
+
 static void *virtqueue_get_buf_ctx_split(struct virtqueue *_vq,
 					 unsigned int *len,
 					 void **ctx)
@@ -919,13 +925,6 @@ static unsigned int virtqueue_enable_cb_prepare_split(struct virtqueue *_vq)
 	return last_used_idx;
 }
 
-static bool virtqueue_poll_split(struct vring_virtqueue *vq,
-				 unsigned int last_used_idx)
-{
-	return (u16)last_used_idx != virtio16_to_cpu(vq->vq.vdev,
-			vq->split.vring.used->idx);
-}
-
 static bool virtqueue_enable_cb_delayed_split(struct virtqueue *_vq)
 {
 	struct vring_virtqueue *vq = to_vvq(_vq);
@@ -1710,16 +1709,20 @@ static inline bool is_used_desc_packed(const struct vring_virtqueue *vq,
 	return avail == used && used == used_wrap_counter;
 }
 
-static bool more_used_packed(const struct vring_virtqueue *vq)
+static bool virtqueue_poll_packed(const struct vring_virtqueue *vq, u16 off_wrap)
 {
-	u16 last_used;
-	u16 last_used_idx;
-	bool used_wrap_counter;
+	bool wrap_counter;
+	u16 used_idx;
 
-	last_used_idx = READ_ONCE(vq->last_used_idx);
-	last_used = packed_last_used(last_used_idx);
-	used_wrap_counter = packed_used_wrap_counter(last_used_idx);
-	return is_used_desc_packed(vq, last_used, used_wrap_counter);
+	wrap_counter = off_wrap >> VRING_PACKED_EVENT_F_WRAP_CTR;
+	used_idx = off_wrap & ~(1 << VRING_PACKED_EVENT_F_WRAP_CTR);
+
+	return is_used_desc_packed(vq, used_idx, wrap_counter);
+}
+
+static bool more_used_packed(const struct vring_virtqueue *vq)
+{
+	return virtqueue_poll_packed(vq, READ_ONCE(vq->last_used_idx));
 }
 
 static void *virtqueue_get_buf_ctx_packed(struct virtqueue *_vq,
@@ -1843,17 +1846,6 @@ static unsigned int virtqueue_enable_cb_prepare_packed(struct virtqueue *_vq)
 	return vq->last_used_idx;
 }
 
-static bool virtqueue_poll_packed(struct vring_virtqueue *vq, u16 off_wrap)
-{
-	bool wrap_counter;
-	u16 used_idx;
-
-	wrap_counter = off_wrap >> VRING_PACKED_EVENT_F_WRAP_CTR;
-	used_idx = off_wrap & ~(1 << VRING_PACKED_EVENT_F_WRAP_CTR);
-
-	return is_used_desc_packed(vq, used_idx, wrap_counter);
-}
-
 static bool virtqueue_enable_cb_delayed_packed(struct virtqueue *_vq)
 {
 	struct vring_virtqueue *vq = to_vvq(_vq);
-- 
2.31.1

[PATCH V9 04/19] virtio_ring: switch to use vring_virtqueue for virtqueue resize variants

[Jason Wang]

Those variants are used internally so let's switch to use
vring_virtqueue as parameter to be consistent with other internal
virtqueue helpers.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 14 ++++++--------
 1 file changed, 6 insertions(+), 8 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index a32050e1aa47..3fe4c2f63dfa 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -1221,11 +1221,10 @@ static struct virtqueue *vring_create_virtqueue_split(
 	return vq;
 }
 
-static int virtqueue_resize_split(struct virtqueue *_vq, u32 num)
+static int virtqueue_resize_split(struct vring_virtqueue *vq, u32 num)
 {
 	struct vring_virtqueue_split vring_split = {};
-	struct vring_virtqueue *vq = to_vvq(_vq);
-	struct virtio_device *vdev = _vq->vdev;
+	struct virtio_device *vdev = vq->vq.vdev;
 	int err;
 
 	err = vring_alloc_queue_split(&vring_split, vdev, num,
@@ -2182,11 +2181,10 @@ static struct virtqueue *vring_create_virtqueue_packed(
 	return vq;
 }
 
-static int virtqueue_resize_packed(struct virtqueue *_vq, u32 num)
+static int virtqueue_resize_packed(struct vring_virtqueue *vq, u32 num)
 {
 	struct vring_virtqueue_packed vring_packed = {};
-	struct vring_virtqueue *vq = to_vvq(_vq);
-	struct virtio_device *vdev = _vq->vdev;
+	struct virtio_device *vdev = vq->vq.vdev;
 	int err;
 
 	if (vring_alloc_queue_packed(&vring_packed, vdev, num, vq->map))
@@ -2809,9 +2807,9 @@ int virtqueue_resize(struct virtqueue *_vq, u32 num,
 		recycle_done(_vq);
 
 	if (vq->packed_ring)
-		err = virtqueue_resize_packed(_vq, num);
+		err = virtqueue_resize_packed(vq, num);
 	else
-		err = virtqueue_resize_split(_vq, num);
+		err = virtqueue_resize_split(vq, num);
 
 	err_reset = virtqueue_enable_after_reset(_vq);
 	if (err_reset)
-- 
2.31.1

[PATCH V9 05/19] virtio_ring: switch to use vring_virtqueue for virtqueue_kick_prepare variants

[Jason Wang]

Those variants are used internally so let's switch to use
vring_virtqueue as parameter to be consistent with other internal
virtqueue helpers.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 14 ++++++--------
 1 file changed, 6 insertions(+), 8 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index 3fe4c2f63dfa..d61d71b12776 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -717,9 +717,8 @@ static inline int virtqueue_add_split(struct virtqueue *_vq,
 	return -ENOMEM;
 }
 
-static bool virtqueue_kick_prepare_split(struct virtqueue *_vq)
+static bool virtqueue_kick_prepare_split(struct vring_virtqueue *vq)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
 	u16 new, old;
 	bool needs_kick;
 
@@ -736,12 +735,12 @@ static bool virtqueue_kick_prepare_split(struct virtqueue *_vq)
 	LAST_ADD_TIME_INVALID(vq);
 
 	if (vq->event) {
-		needs_kick = vring_need_event(virtio16_to_cpu(_vq->vdev,
+		needs_kick = vring_need_event(virtio16_to_cpu(vq->vq.vdev,
 					vring_avail_event(&vq->split.vring)),
 					      new, old);
 	} else {
 		needs_kick = !(vq->split.vring.used->flags &
-					cpu_to_virtio16(_vq->vdev,
+					cpu_to_virtio16(vq->vq.vdev,
 						VRING_USED_F_NO_NOTIFY));
 	}
 	END_USE(vq);
@@ -1596,9 +1595,8 @@ static inline int virtqueue_add_packed(struct virtqueue *_vq,
 	return -EIO;
 }
 
-static bool virtqueue_kick_prepare_packed(struct virtqueue *_vq)
+static bool virtqueue_kick_prepare_packed(struct vring_virtqueue *vq)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
 	u16 new, old, off_wrap, flags, wrap_counter, event_idx;
 	bool needs_kick;
 	union {
@@ -2457,8 +2455,8 @@ bool virtqueue_kick_prepare(struct virtqueue *_vq)
 {
 	struct vring_virtqueue *vq = to_vvq(_vq);
 
-	return vq->packed_ring ? virtqueue_kick_prepare_packed(_vq) :
-				 virtqueue_kick_prepare_split(_vq);
+	return vq->packed_ring ? virtqueue_kick_prepare_packed(vq) :
+				 virtqueue_kick_prepare_split(vq);
 }
 EXPORT_SYMBOL_GPL(virtqueue_kick_prepare);
 
-- 
2.31.1

[PATCH V9 06/19] virtio_ring: switch to use vring_virtqueue for virtqueue_add variants

[Jason Wang]

Those variants are used internally so let's switch to use
vring_virtqueue as parameter to be consistent with other internal
virtqueue helpers.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 39 ++++++++++++++++++------------------
 1 file changed, 19 insertions(+), 20 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index d61d71b12776..f200d812416e 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -476,7 +476,7 @@ static unsigned int vring_unmap_one_split(const struct vring_virtqueue *vq,
 	return extra->next;
 }
 
-static struct vring_desc *alloc_indirect_split(struct virtqueue *_vq,
+static struct vring_desc *alloc_indirect_split(struct vring_virtqueue *vq,
 					       unsigned int total_sg,
 					       gfp_t gfp)
 {
@@ -505,7 +505,7 @@ static struct vring_desc *alloc_indirect_split(struct virtqueue *_vq,
 	return desc;
 }
 
-static inline unsigned int virtqueue_add_desc_split(struct virtqueue *vq,
+static inline unsigned int virtqueue_add_desc_split(struct vring_virtqueue *vq,
 						    struct vring_desc *desc,
 						    struct vring_desc_extra *extra,
 						    unsigned int i,
@@ -513,11 +513,12 @@ static inline unsigned int virtqueue_add_desc_split(struct virtqueue *vq,
 						    unsigned int len,
 						    u16 flags, bool premapped)
 {
+	struct virtio_device *vdev = vq->vq.vdev;
 	u16 next;
 
-	desc[i].flags = cpu_to_virtio16(vq->vdev, flags);
-	desc[i].addr = cpu_to_virtio64(vq->vdev, addr);
-	desc[i].len = cpu_to_virtio32(vq->vdev, len);
+	desc[i].flags = cpu_to_virtio16(vdev, flags);
+	desc[i].addr = cpu_to_virtio64(vdev, addr);
+	desc[i].len = cpu_to_virtio32(vdev, len);
 
 	extra[i].addr = premapped ? DMA_MAPPING_ERROR : addr;
 	extra[i].len = len;
@@ -525,12 +526,12 @@ static inline unsigned int virtqueue_add_desc_split(struct virtqueue *vq,
 
 	next = extra[i].next;
 
-	desc[i].next = cpu_to_virtio16(vq->vdev, next);
+	desc[i].next = cpu_to_virtio16(vdev, next);
 
 	return next;
 }
 
-static inline int virtqueue_add_split(struct virtqueue *_vq,
+static inline int virtqueue_add_split(struct vring_virtqueue *vq,
 				      struct scatterlist *sgs[],
 				      unsigned int total_sg,
 				      unsigned int out_sgs,
@@ -540,7 +541,6 @@ static inline int virtqueue_add_split(struct virtqueue *_vq,
 				      bool premapped,
 				      gfp_t gfp)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
 	struct vring_desc_extra *extra;
 	struct scatterlist *sg;
 	struct vring_desc *desc;
@@ -565,7 +565,7 @@ static inline int virtqueue_add_split(struct virtqueue *_vq,
 	head = vq->free_head;
 
 	if (virtqueue_use_indirect(vq, total_sg))
-		desc = alloc_indirect_split(_vq, total_sg, gfp);
+		desc = alloc_indirect_split(vq, total_sg, gfp);
 	else {
 		desc = NULL;
 		WARN_ON_ONCE(total_sg > vq->split.vring.num && !vq->indirect);
@@ -612,7 +612,7 @@ static inline int virtqueue_add_split(struct virtqueue *_vq,
 			/* Note that we trust indirect descriptor
 			 * table since it use stream DMA mapping.
 			 */
-			i = virtqueue_add_desc_split(_vq, desc, extra, i, addr, len,
+			i = virtqueue_add_desc_split(vq, desc, extra, i, addr, len,
 						     VRING_DESC_F_NEXT,
 						     premapped);
 		}
@@ -629,14 +629,14 @@ static inline int virtqueue_add_split(struct virtqueue *_vq,
 			/* Note that we trust indirect descriptor
 			 * table since it use stream DMA mapping.
 			 */
-			i = virtqueue_add_desc_split(_vq, desc, extra, i, addr, len,
+			i = virtqueue_add_desc_split(vq, desc, extra, i, addr, len,
 						     VRING_DESC_F_NEXT |
 						     VRING_DESC_F_WRITE,
 						     premapped);
 		}
 	}
 	/* Last one doesn't continue. */
-	desc[prev].flags &= cpu_to_virtio16(_vq->vdev, ~VRING_DESC_F_NEXT);
+	desc[prev].flags &= cpu_to_virtio16(vq->vq.vdev, ~VRING_DESC_F_NEXT);
 	if (!indirect && vring_need_unmap_buffer(vq, &extra[prev]))
 		vq->split.desc_extra[prev & (vq->split.vring.num - 1)].flags &=
 			~VRING_DESC_F_NEXT;
@@ -649,7 +649,7 @@ static inline int virtqueue_add_split(struct virtqueue *_vq,
 		if (vring_mapping_error(vq, addr))
 			goto unmap_release;
 
-		virtqueue_add_desc_split(_vq, vq->split.vring.desc,
+		virtqueue_add_desc_split(vq, vq->split.vring.desc,
 					 vq->split.desc_extra,
 					 head, addr,
 					 total_sg * sizeof(struct vring_desc),
@@ -675,13 +675,13 @@ static inline int virtqueue_add_split(struct virtqueue *_vq,
 	/* Put entry in available array (but don't update avail->idx until they
 	 * do sync). */
 	avail = vq->split.avail_idx_shadow & (vq->split.vring.num - 1);
-	vq->split.vring.avail->ring[avail] = cpu_to_virtio16(_vq->vdev, head);
+	vq->split.vring.avail->ring[avail] = cpu_to_virtio16(vq->vq.vdev, head);
 
 	/* Descriptors and available array need to be set before we expose the
 	 * new available array entries. */
 	virtio_wmb(vq->weak_barriers);
 	vq->split.avail_idx_shadow++;
-	vq->split.vring.avail->idx = cpu_to_virtio16(_vq->vdev,
+	vq->split.vring.avail->idx = cpu_to_virtio16(vq->vq.vdev,
 						vq->split.avail_idx_shadow);
 	vq->num_added++;
 
@@ -691,7 +691,7 @@ static inline int virtqueue_add_split(struct virtqueue *_vq,
 	/* This is very unlikely, but theoretically possible.  Kick
 	 * just in case. */
 	if (unlikely(vq->num_added == (1 << 16) - 1))
-		virtqueue_kick(_vq);
+		virtqueue_kick(&vq->vq);
 
 	return 0;
 
@@ -1440,7 +1440,7 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
 	return -ENOMEM;
 }
 
-static inline int virtqueue_add_packed(struct virtqueue *_vq,
+static inline int virtqueue_add_packed(struct vring_virtqueue *vq,
 				       struct scatterlist *sgs[],
 				       unsigned int total_sg,
 				       unsigned int out_sgs,
@@ -1450,7 +1450,6 @@ static inline int virtqueue_add_packed(struct virtqueue *_vq,
 				       bool premapped,
 				       gfp_t gfp)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
 	struct vring_packed_desc *desc;
 	struct scatterlist *sg;
 	unsigned int i, n, c, descs_used, err_idx, len;
@@ -2262,9 +2261,9 @@ static inline int virtqueue_add(struct virtqueue *_vq,
 {
 	struct vring_virtqueue *vq = to_vvq(_vq);
 
-	return vq->packed_ring ? virtqueue_add_packed(_vq, sgs, total_sg,
+	return vq->packed_ring ? virtqueue_add_packed(vq, sgs, total_sg,
 					out_sgs, in_sgs, data, ctx, premapped, gfp) :
-				 virtqueue_add_split(_vq, sgs, total_sg,
+				 virtqueue_add_split(vq, sgs, total_sg,
 					out_sgs, in_sgs, data, ctx, premapped, gfp);
 }
 
-- 
2.31.1

[PATCH V9 07/19] virtio: switch to use vring_virtqueue for virtqueue_get variants

[Jason Wang]

Those variants are used internally so let's switch to use
vring_virtqueue as parameter to be consistent with other internal
virtqueue helpers.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 16 +++++++---------
 1 file changed, 7 insertions(+), 9 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index f200d812416e..3fd65b1177a3 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -817,11 +817,10 @@ static bool more_used_split(const struct vring_virtqueue *vq)
 	return virtqueue_poll_split(vq, vq->last_used_idx);
 }
 
-static void *virtqueue_get_buf_ctx_split(struct virtqueue *_vq,
+static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
 					 unsigned int *len,
 					 void **ctx)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
 	void *ret;
 	unsigned int i;
 	u16 last_used;
@@ -843,9 +842,9 @@ static void *virtqueue_get_buf_ctx_split(struct virtqueue *_vq,
 	virtio_rmb(vq->weak_barriers);
 
 	last_used = (vq->last_used_idx & (vq->split.vring.num - 1));
-	i = virtio32_to_cpu(_vq->vdev,
+	i = virtio32_to_cpu(vq->vq.vdev,
 			vq->split.vring.used->ring[last_used].id);
-	*len = virtio32_to_cpu(_vq->vdev,
+	*len = virtio32_to_cpu(vq->vq.vdev,
 			vq->split.vring.used->ring[last_used].len);
 
 	if (unlikely(i >= vq->split.vring.num)) {
@@ -867,7 +866,7 @@ static void *virtqueue_get_buf_ctx_split(struct virtqueue *_vq,
 	if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
 		virtio_store_mb(vq->weak_barriers,
 				&vring_used_event(&vq->split.vring),
-				cpu_to_virtio16(_vq->vdev, vq->last_used_idx));
+				cpu_to_virtio16(vq->vq.vdev, vq->last_used_idx));
 
 	LAST_ADD_TIME_INVALID(vq);
 
@@ -1721,11 +1720,10 @@ static bool more_used_packed(const struct vring_virtqueue *vq)
 	return virtqueue_poll_packed(vq, READ_ONCE(vq->last_used_idx));
 }
 
-static void *virtqueue_get_buf_ctx_packed(struct virtqueue *_vq,
+static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
 					  unsigned int *len,
 					  void **ctx)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
 	u16 last_used, id, last_used_idx;
 	bool used_wrap_counter;
 	void *ret;
@@ -2525,8 +2523,8 @@ void *virtqueue_get_buf_ctx(struct virtqueue *_vq, unsigned int *len,
 {
 	struct vring_virtqueue *vq = to_vvq(_vq);
 
-	return vq->packed_ring ? virtqueue_get_buf_ctx_packed(_vq, len, ctx) :
-				 virtqueue_get_buf_ctx_split(_vq, len, ctx);
+	return vq->packed_ring ? virtqueue_get_buf_ctx_packed(vq, len, ctx) :
+				 virtqueue_get_buf_ctx_split(vq, len, ctx);
 }
 EXPORT_SYMBOL_GPL(virtqueue_get_buf_ctx);
 
-- 
2.31.1

[PATCH V9 08/19] virtio_ring: switch to use vring_virtqueue for enable_cb_prepare variants

[Jason Wang]

Those variants are used internally so let's switch to use
vring_virtqueue as parameter to be consistent with other internal
virtqueue helpers.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 15 ++++++---------
 1 file changed, 6 insertions(+), 9 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index 3fd65b1177a3..8ffe7daa1e5f 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -898,9 +898,8 @@ static void virtqueue_disable_cb_split(struct virtqueue *_vq)
 	}
 }
 
-static unsigned int virtqueue_enable_cb_prepare_split(struct virtqueue *_vq)
+static unsigned int virtqueue_enable_cb_prepare_split(struct vring_virtqueue *vq)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
 	u16 last_used_idx;
 
 	START_USE(vq);
@@ -914,10 +913,10 @@ static unsigned int virtqueue_enable_cb_prepare_split(struct virtqueue *_vq)
 		vq->split.avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
 		if (!vq->event)
 			vq->split.vring.avail->flags =
-				cpu_to_virtio16(_vq->vdev,
+				cpu_to_virtio16(vq->vq.vdev,
 						vq->split.avail_flags_shadow);
 	}
-	vring_used_event(&vq->split.vring) = cpu_to_virtio16(_vq->vdev,
+	vring_used_event(&vq->split.vring) = cpu_to_virtio16(vq->vq.vdev,
 			last_used_idx = vq->last_used_idx);
 	END_USE(vq);
 	return last_used_idx;
@@ -1807,10 +1806,8 @@ static void virtqueue_disable_cb_packed(struct virtqueue *_vq)
 	}
 }
 
-static unsigned int virtqueue_enable_cb_prepare_packed(struct virtqueue *_vq)
+static unsigned int virtqueue_enable_cb_prepare_packed(struct vring_virtqueue *vq)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
-
 	START_USE(vq);
 
 	/*
@@ -2572,8 +2569,8 @@ unsigned int virtqueue_enable_cb_prepare(struct virtqueue *_vq)
 	if (vq->event_triggered)
 		vq->event_triggered = false;
 
-	return vq->packed_ring ? virtqueue_enable_cb_prepare_packed(_vq) :
-				 virtqueue_enable_cb_prepare_split(_vq);
+	return vq->packed_ring ? virtqueue_enable_cb_prepare_packed(vq) :
+				 virtqueue_enable_cb_prepare_split(vq);
 }
 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_prepare);
 
-- 
2.31.1

[PATCH V9 09/19] virtio_ring: use vring_virtqueue for enable_cb_delayed variants

[Jason Wang]

Those variants are used internally so let's switch to use
vring_virtqueue as parameter to be consistent with other internal
virtqueue helpers.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 16 +++++++---------
 1 file changed, 7 insertions(+), 9 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index 8ffe7daa1e5f..00de30862777 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -922,9 +922,8 @@ static unsigned int virtqueue_enable_cb_prepare_split(struct vring_virtqueue *vq
 	return last_used_idx;
 }
 
-static bool virtqueue_enable_cb_delayed_split(struct virtqueue *_vq)
+static bool virtqueue_enable_cb_delayed_split(struct vring_virtqueue *vq)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
 	u16 bufs;
 
 	START_USE(vq);
@@ -938,7 +937,7 @@ static bool virtqueue_enable_cb_delayed_split(struct virtqueue *_vq)
 		vq->split.avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
 		if (!vq->event)
 			vq->split.vring.avail->flags =
-				cpu_to_virtio16(_vq->vdev,
+				cpu_to_virtio16(vq->vq.vdev,
 						vq->split.avail_flags_shadow);
 	}
 	/* TODO: tune this threshold */
@@ -946,9 +945,9 @@ static bool virtqueue_enable_cb_delayed_split(struct virtqueue *_vq)
 
 	virtio_store_mb(vq->weak_barriers,
 			&vring_used_event(&vq->split.vring),
-			cpu_to_virtio16(_vq->vdev, vq->last_used_idx + bufs));
+			cpu_to_virtio16(vq->vq.vdev, vq->last_used_idx + bufs));
 
-	if (unlikely((u16)(virtio16_to_cpu(_vq->vdev, vq->split.vring.used->idx)
+	if (unlikely((u16)(virtio16_to_cpu(vq->vq.vdev, vq->split.vring.used->idx)
 					- vq->last_used_idx) > bufs)) {
 		END_USE(vq);
 		return false;
@@ -1837,9 +1836,8 @@ static unsigned int virtqueue_enable_cb_prepare_packed(struct vring_virtqueue *v
 	return vq->last_used_idx;
 }
 
-static bool virtqueue_enable_cb_delayed_packed(struct virtqueue *_vq)
+static bool virtqueue_enable_cb_delayed_packed(struct vring_virtqueue *vq)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
 	u16 used_idx, wrap_counter, last_used_idx;
 	u16 bufs;
 
@@ -2635,8 +2633,8 @@ bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
 	if (vq->event_triggered)
 		data_race(vq->event_triggered = false);
 
-	return vq->packed_ring ? virtqueue_enable_cb_delayed_packed(_vq) :
-				 virtqueue_enable_cb_delayed_split(_vq);
+	return vq->packed_ring ? virtqueue_enable_cb_delayed_packed(vq) :
+				 virtqueue_enable_cb_delayed_split(vq);
 }
 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
 
-- 
2.31.1

[PATCH V9 10/19] virtio_ring: switch to use vring_virtqueue for disable_cb variants

[Jason Wang]

Those variants are used internally so let's switch to use
vring_virtqueue as parameter to be consistent with other internal
virtqueue helpers.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 14 +++++---------
 1 file changed, 5 insertions(+), 9 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index 00de30862777..a4d084f5ca95 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -874,10 +874,8 @@ static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
 	return ret;
 }
 
-static void virtqueue_disable_cb_split(struct virtqueue *_vq)
+static void virtqueue_disable_cb_split(struct vring_virtqueue *vq)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
-
 	if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
 		vq->split.avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
 
@@ -893,7 +891,7 @@ static void virtqueue_disable_cb_split(struct virtqueue *_vq)
 			vring_used_event(&vq->split.vring) = 0x0;
 		else
 			vq->split.vring.avail->flags =
-				cpu_to_virtio16(_vq->vdev,
+				cpu_to_virtio16(vq->vq.vdev,
 						vq->split.avail_flags_shadow);
 	}
 }
@@ -1786,10 +1784,8 @@ static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
 	return ret;
 }
 
-static void virtqueue_disable_cb_packed(struct virtqueue *_vq)
+static void virtqueue_disable_cb_packed(struct vring_virtqueue *vq)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
-
 	if (vq->packed.event_flags_shadow != VRING_PACKED_EVENT_FLAG_DISABLE) {
 		vq->packed.event_flags_shadow = VRING_PACKED_EVENT_FLAG_DISABLE;
 
@@ -2542,9 +2538,9 @@ void virtqueue_disable_cb(struct virtqueue *_vq)
 	struct vring_virtqueue *vq = to_vvq(_vq);
 
 	if (vq->packed_ring)
-		virtqueue_disable_cb_packed(_vq);
+		virtqueue_disable_cb_packed(vq);
 	else
-		virtqueue_disable_cb_split(_vq);
+		virtqueue_disable_cb_split(vq);
 }
 EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
 
-- 
2.31.1

[PATCH V9 11/19] virtio_ring: switch to use vring_virtqueue for detach_unused_buf variants

[Jason Wang]

Those variants are used internally so let's switch to use
vring_virtqueue as parameter to be consistent with other internal
virtqueue helpers.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 12 +++++-------
 1 file changed, 5 insertions(+), 7 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index a4d084f5ca95..db5cdb5fcb20 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -955,9 +955,8 @@ static bool virtqueue_enable_cb_delayed_split(struct vring_virtqueue *vq)
 	return true;
 }
 
-static void *virtqueue_detach_unused_buf_split(struct virtqueue *_vq)
+static void *virtqueue_detach_unused_buf_split(struct vring_virtqueue *vq)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
 	unsigned int i;
 	void *buf;
 
@@ -970,7 +969,7 @@ static void *virtqueue_detach_unused_buf_split(struct virtqueue *_vq)
 		buf = vq->split.desc_state[i].data;
 		detach_buf_split(vq, i, NULL);
 		vq->split.avail_idx_shadow--;
-		vq->split.vring.avail->idx = cpu_to_virtio16(_vq->vdev,
+		vq->split.vring.avail->idx = cpu_to_virtio16(vq->vq.vdev,
 				vq->split.avail_idx_shadow);
 		END_USE(vq);
 		return buf;
@@ -1892,9 +1891,8 @@ static bool virtqueue_enable_cb_delayed_packed(struct vring_virtqueue *vq)
 	return true;
 }
 
-static void *virtqueue_detach_unused_buf_packed(struct virtqueue *_vq)
+static void *virtqueue_detach_unused_buf_packed(struct vring_virtqueue *vq)
 {
-	struct vring_virtqueue *vq = to_vvq(_vq);
 	unsigned int i;
 	void *buf;
 
@@ -2646,8 +2644,8 @@ void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
 {
 	struct vring_virtqueue *vq = to_vvq(_vq);
 
-	return vq->packed_ring ? virtqueue_detach_unused_buf_packed(_vq) :
-				 virtqueue_detach_unused_buf_split(_vq);
+	return vq->packed_ring ? virtqueue_detach_unused_buf_packed(vq) :
+				 virtqueue_detach_unused_buf_split(vq);
 }
 EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
 
-- 
2.31.1

[PATCH V9 12/19] virtio_ring: switch to use unsigned int for virtqueue_poll_packed()

[Jason Wang]

Switch to use unsigned int for virtqueue_poll_packed() to match
virtqueue_poll() and virtqueue_poll_split() and to ease the
abstraction of the virtqueue ops.

Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index db5cdb5fcb20..f804c5c6975a 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -1699,7 +1699,8 @@ static inline bool is_used_desc_packed(const struct vring_virtqueue *vq,
 	return avail == used && used == used_wrap_counter;
 }
 
-static bool virtqueue_poll_packed(const struct vring_virtqueue *vq, u16 off_wrap)
+static bool virtqueue_poll_packed(const struct vring_virtqueue *vq,
+				  unsigned int off_wrap)
 {
 	bool wrap_counter;
 	u16 used_idx;
-- 
2.31.1

[PATCH V9 13/19] virtio_ring: introduce virtqueue ops

[Jason Wang]

This patch introduces virtqueue ops which is a set of callbacks
that will be called for different queue layout or features. This would
help to avoid branches for split/packed and will ease the future
implementation like in order.

Note that in order to eliminate the indirect calls this patch uses
global array of const ops to allow compiler to avoid indirect
branches.

Tested with CONFIG_MITIGATION_RETPOLINE, no performance differences
were noticed.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Suggested-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 173 ++++++++++++++++++++++++++---------
 1 file changed, 130 insertions(+), 43 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index f804c5c6975a..4e905dc7b6cb 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -67,6 +67,12 @@
 #define LAST_ADD_TIME_INVALID(vq)
 #endif
 
+enum vq_layout {
+	SPLIT = 0,
+	PACKED,
+	VQ_TYPE_MAX,
+};
+
 struct vring_desc_state_split {
 	void *data;			/* Data for callback. */
 
@@ -159,12 +165,29 @@ struct vring_virtqueue_packed {
 	size_t event_size_in_bytes;
 };
 
+struct vring_virtqueue;
+
+struct virtqueue_ops {
+	int (*add)(struct vring_virtqueue *vq, struct scatterlist *sgs[],
+		   unsigned int total_sg, unsigned int out_sgs,
+		   unsigned int in_sgs,	void *data,
+		   void *ctx, bool premapped, gfp_t gfp);
+	void *(*get)(struct vring_virtqueue *vq, unsigned int *len, void **ctx);
+	bool (*kick_prepare)(struct vring_virtqueue *vq);
+	void (*disable_cb)(struct vring_virtqueue *vq);
+	bool (*enable_cb_delayed)(struct vring_virtqueue *vq);
+	unsigned int (*enable_cb_prepare)(struct vring_virtqueue *vq);
+	bool (*poll)(const struct vring_virtqueue *vq,
+		     unsigned int last_used_idx);
+	void *(*detach_unused_buf)(struct vring_virtqueue *vq);
+	bool (*more_used)(const struct vring_virtqueue *vq);
+	int (*resize)(struct vring_virtqueue *vq, u32 num);
+	void (*reset)(struct vring_virtqueue *vq);
+};
+
 struct vring_virtqueue {
 	struct virtqueue vq;
 
-	/* Is this a packed ring? */
-	bool packed_ring;
-
 	/* Is DMA API used? */
 	bool use_map_api;
 
@@ -180,6 +203,8 @@ struct vring_virtqueue {
 	/* Host publishes avail event idx */
 	bool event;
 
+	enum vq_layout layout;
+
 	/* Head of free buffer list. */
 	unsigned int free_head;
 	/* Number we've added since last sync. */
@@ -231,6 +256,12 @@ static void vring_free(struct virtqueue *_vq);
 
 #define to_vvq(_vq) container_of_const(_vq, struct vring_virtqueue, vq)
 
+
+static inline bool virtqueue_is_packed(const struct vring_virtqueue *vq)
+{
+	return vq->layout == PACKED;
+}
+
 static bool virtqueue_use_indirect(const struct vring_virtqueue *vq,
 				   unsigned int total_sg)
 {
@@ -433,7 +464,7 @@ static void virtqueue_init(struct vring_virtqueue *vq, u32 num)
 {
 	vq->vq.num_free = num;
 
-	if (vq->packed_ring)
+	if (virtqueue_is_packed(vq))
 		vq->last_used_idx = 0 | (1 << VRING_PACKED_EVENT_F_WRAP_CTR);
 	else
 		vq->last_used_idx = 0;
@@ -1122,6 +1153,8 @@ static int vring_alloc_queue_split(struct vring_virtqueue_split *vring_split,
 	return 0;
 }
 
+static const struct virtqueue_ops split_ops;
+
 static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
 					       struct vring_virtqueue_split *vring_split,
 					       struct virtio_device *vdev,
@@ -1139,7 +1172,7 @@ static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
 	if (!vq)
 		return NULL;
 
-	vq->packed_ring = false;
+	vq->layout = SPLIT;
 	vq->vq.callback = callback;
 	vq->vq.vdev = vdev;
 	vq->vq.name = name;
@@ -2077,6 +2110,8 @@ static void virtqueue_reset_packed(struct vring_virtqueue *vq)
 	virtqueue_vring_init_packed(&vq->packed, !!vq->vq.callback);
 }
 
+static const struct virtqueue_ops packed_ops;
+
 static struct virtqueue *__vring_new_virtqueue_packed(unsigned int index,
 					       struct vring_virtqueue_packed *vring_packed,
 					       struct virtio_device *vdev,
@@ -2107,7 +2142,7 @@ static struct virtqueue *__vring_new_virtqueue_packed(unsigned int index,
 #else
 	vq->broken = false;
 #endif
-	vq->packed_ring = true;
+	vq->layout = PACKED;
 	vq->map = map;
 	vq->use_map_api = vring_use_map_api(vdev);
 
@@ -2195,6 +2230,39 @@ static int virtqueue_resize_packed(struct vring_virtqueue *vq, u32 num)
 	return -ENOMEM;
 }
 
+static const struct virtqueue_ops split_ops = {
+	.add = virtqueue_add_split,
+	.get = virtqueue_get_buf_ctx_split,
+	.kick_prepare = virtqueue_kick_prepare_split,
+	.disable_cb = virtqueue_disable_cb_split,
+	.enable_cb_delayed = virtqueue_enable_cb_delayed_split,
+	.enable_cb_prepare = virtqueue_enable_cb_prepare_split,
+	.poll = virtqueue_poll_split,
+	.detach_unused_buf = virtqueue_detach_unused_buf_split,
+	.more_used = more_used_split,
+	.resize = virtqueue_resize_split,
+	.reset = virtqueue_reset_split,
+};
+
+static const struct virtqueue_ops packed_ops = {
+	.add = virtqueue_add_packed,
+	.get = virtqueue_get_buf_ctx_packed,
+	.kick_prepare = virtqueue_kick_prepare_packed,
+	.disable_cb = virtqueue_disable_cb_packed,
+	.enable_cb_delayed = virtqueue_enable_cb_delayed_packed,
+	.enable_cb_prepare = virtqueue_enable_cb_prepare_packed,
+	.poll = virtqueue_poll_packed,
+	.detach_unused_buf = virtqueue_detach_unused_buf_packed,
+	.more_used = more_used_packed,
+	.resize = virtqueue_resize_packed,
+	.reset = virtqueue_reset_packed,
+};
+
+static const struct virtqueue_ops *const all_ops[VQ_TYPE_MAX] = {
+	[SPLIT] = &split_ops,
+	[PACKED] = &packed_ops
+};
+
 static int virtqueue_disable_and_recycle(struct virtqueue *_vq,
 					 void (*recycle)(struct virtqueue *vq, void *buf))
 {
@@ -2237,6 +2305,39 @@ static int virtqueue_enable_after_reset(struct virtqueue *_vq)
  * Generic functions and exported symbols.
  */
 
+#define VIRTQUEUE_CALL(vq, op, ...)					\
+	({								\
+	typeof(all_ops[SPLIT]->op(vq, ##__VA_ARGS__)) ret;		\
+									\
+	switch (vq->layout) {						\
+	case SPLIT:							\
+		ret = all_ops[SPLIT]->op(vq, ##__VA_ARGS__);		\
+		break;							\
+	case PACKED:							\
+		ret = all_ops[PACKED]->op(vq, ##__VA_ARGS__);		\
+		break;							\
+	default:							\
+		BUG();							\
+		break;							\
+	}								\
+	ret;								\
+})
+
+#define VOID_VIRTQUEUE_CALL(vq, op, ...)		\
+	({						\
+	switch ((vq)->layout) {			\
+	case SPLIT:					\
+		all_ops[SPLIT]->op(vq, ##__VA_ARGS__);	\
+		break;					\
+	case PACKED:					\
+		all_ops[PACKED]->op(vq, ##__VA_ARGS__);	\
+		break;					\
+	default:					\
+		BUG();					\
+		break;					\
+	}						\
+})
+
 static inline int virtqueue_add(struct virtqueue *_vq,
 				struct scatterlist *sgs[],
 				unsigned int total_sg,
@@ -2249,10 +2350,9 @@ static inline int virtqueue_add(struct virtqueue *_vq,
 {
 	struct vring_virtqueue *vq = to_vvq(_vq);
 
-	return vq->packed_ring ? virtqueue_add_packed(vq, sgs, total_sg,
-					out_sgs, in_sgs, data, ctx, premapped, gfp) :
-				 virtqueue_add_split(vq, sgs, total_sg,
-					out_sgs, in_sgs, data, ctx, premapped, gfp);
+	return VIRTQUEUE_CALL(vq, add, sgs, total_sg,
+			      out_sgs, in_sgs, data,
+			      ctx, premapped, gfp);
 }
 
 /**
@@ -2442,8 +2542,7 @@ bool virtqueue_kick_prepare(struct virtqueue *_vq)
 {
 	struct vring_virtqueue *vq = to_vvq(_vq);
 
-	return vq->packed_ring ? virtqueue_kick_prepare_packed(vq) :
-				 virtqueue_kick_prepare_split(vq);
+	return VIRTQUEUE_CALL(vq, kick_prepare);
 }
 EXPORT_SYMBOL_GPL(virtqueue_kick_prepare);
 
@@ -2513,8 +2612,7 @@ void *virtqueue_get_buf_ctx(struct virtqueue *_vq, unsigned int *len,
 {
 	struct vring_virtqueue *vq = to_vvq(_vq);
 
-	return vq->packed_ring ? virtqueue_get_buf_ctx_packed(vq, len, ctx) :
-				 virtqueue_get_buf_ctx_split(vq, len, ctx);
+	return VIRTQUEUE_CALL(vq, get, len, ctx);
 }
 EXPORT_SYMBOL_GPL(virtqueue_get_buf_ctx);
 
@@ -2536,10 +2634,7 @@ void virtqueue_disable_cb(struct virtqueue *_vq)
 {
 	struct vring_virtqueue *vq = to_vvq(_vq);
 
-	if (vq->packed_ring)
-		virtqueue_disable_cb_packed(vq);
-	else
-		virtqueue_disable_cb_split(vq);
+	VOID_VIRTQUEUE_CALL(vq, disable_cb);
 }
 EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
 
@@ -2562,8 +2657,7 @@ unsigned int virtqueue_enable_cb_prepare(struct virtqueue *_vq)
 	if (vq->event_triggered)
 		vq->event_triggered = false;
 
-	return vq->packed_ring ? virtqueue_enable_cb_prepare_packed(vq) :
-				 virtqueue_enable_cb_prepare_split(vq);
+	return VIRTQUEUE_CALL(vq, enable_cb_prepare);
 }
 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_prepare);
 
@@ -2584,8 +2678,8 @@ bool virtqueue_poll(struct virtqueue *_vq, unsigned int last_used_idx)
 		return false;
 
 	virtio_mb(vq->weak_barriers);
-	return vq->packed_ring ? virtqueue_poll_packed(vq, last_used_idx) :
-				 virtqueue_poll_split(vq, last_used_idx);
+
+	return VIRTQUEUE_CALL(vq, poll, last_used_idx);
 }
 EXPORT_SYMBOL_GPL(virtqueue_poll);
 
@@ -2628,8 +2722,7 @@ bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
 	if (vq->event_triggered)
 		data_race(vq->event_triggered = false);
 
-	return vq->packed_ring ? virtqueue_enable_cb_delayed_packed(vq) :
-				 virtqueue_enable_cb_delayed_split(vq);
+	return VIRTQUEUE_CALL(vq, enable_cb_delayed);
 }
 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
 
@@ -2645,14 +2738,13 @@ void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
 {
 	struct vring_virtqueue *vq = to_vvq(_vq);
 
-	return vq->packed_ring ? virtqueue_detach_unused_buf_packed(vq) :
-				 virtqueue_detach_unused_buf_split(vq);
+	return VIRTQUEUE_CALL(vq, detach_unused_buf);
 }
 EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
 
 static inline bool more_used(const struct vring_virtqueue *vq)
 {
-	return vq->packed_ring ? more_used_packed(vq) : more_used_split(vq);
+	return VIRTQUEUE_CALL(vq, more_used);
 }
 
 /**
@@ -2782,7 +2874,7 @@ int virtqueue_resize(struct virtqueue *_vq, u32 num,
 	if (!num)
 		return -EINVAL;
 
-	if ((vq->packed_ring ? vq->packed.vring.num : vq->split.vring.num) == num)
+	if (virtqueue_get_vring_size(_vq) == num)
 		return 0;
 
 	err = virtqueue_disable_and_recycle(_vq, recycle);
@@ -2791,10 +2883,7 @@ int virtqueue_resize(struct virtqueue *_vq, u32 num,
 	if (recycle_done)
 		recycle_done(_vq);
 
-	if (vq->packed_ring)
-		err = virtqueue_resize_packed(vq, num);
-	else
-		err = virtqueue_resize_split(vq, num);
+	err = VIRTQUEUE_CALL(vq, resize, num);
 
 	err_reset = virtqueue_enable_after_reset(_vq);
 	if (err_reset)
@@ -2832,10 +2921,7 @@ int virtqueue_reset(struct virtqueue *_vq,
 	if (recycle_done)
 		recycle_done(_vq);
 
-	if (vq->packed_ring)
-		virtqueue_reset_packed(vq);
-	else
-		virtqueue_reset_split(vq);
+	VOID_VIRTQUEUE_CALL(vq, reset);
 
 	return virtqueue_enable_after_reset(_vq);
 }
@@ -2878,7 +2964,7 @@ static void vring_free(struct virtqueue *_vq)
 	struct vring_virtqueue *vq = to_vvq(_vq);
 
 	if (vq->we_own_ring) {
-		if (vq->packed_ring) {
+		if (virtqueue_is_packed(vq)) {
 			vring_free_queue(vq->vq.vdev,
 					 vq->packed.ring_size_in_bytes,
 					 vq->packed.vring.desc,
@@ -2907,7 +2993,7 @@ static void vring_free(struct virtqueue *_vq)
 					 vq->map);
 		}
 	}
-	if (!vq->packed_ring) {
+	if (!virtqueue_is_packed(vq)) {
 		kfree(vq->split.desc_state);
 		kfree(vq->split.desc_extra);
 	}
@@ -2932,7 +3018,7 @@ u32 vring_notification_data(struct virtqueue *_vq)
 	struct vring_virtqueue *vq = to_vvq(_vq);
 	u16 next;
 
-	if (vq->packed_ring)
+	if (virtqueue_is_packed(vq))
 		next = (vq->packed.next_avail_idx &
 				~(-(1 << VRING_PACKED_EVENT_F_WRAP_CTR))) |
 			vq->packed.avail_wrap_counter <<
@@ -2985,7 +3071,8 @@ unsigned int virtqueue_get_vring_size(const struct virtqueue *_vq)
 
 	const struct vring_virtqueue *vq = to_vvq(_vq);
 
-	return vq->packed_ring ? vq->packed.vring.num : vq->split.vring.num;
+	return virtqueue_is_packed(vq) ? vq->packed.vring.num :
+				      vq->split.vring.num;
 }
 EXPORT_SYMBOL_GPL(virtqueue_get_vring_size);
 
@@ -3068,7 +3155,7 @@ dma_addr_t virtqueue_get_desc_addr(const struct virtqueue *_vq)
 
 	BUG_ON(!vq->we_own_ring);
 
-	if (vq->packed_ring)
+	if (virtqueue_is_packed(vq))
 		return vq->packed.ring_dma_addr;
 
 	return vq->split.queue_dma_addr;
@@ -3081,7 +3168,7 @@ dma_addr_t virtqueue_get_avail_addr(const struct virtqueue *_vq)
 
 	BUG_ON(!vq->we_own_ring);
 
-	if (vq->packed_ring)
+	if (virtqueue_is_packed(vq))
 		return vq->packed.driver_event_dma_addr;
 
 	return vq->split.queue_dma_addr +
@@ -3095,7 +3182,7 @@ dma_addr_t virtqueue_get_used_addr(const struct virtqueue *_vq)
 
 	BUG_ON(!vq->we_own_ring);
 
-	if (vq->packed_ring)
+	if (virtqueue_is_packed(vq))
 		return vq->packed.device_event_dma_addr;
 
 	return vq->split.queue_dma_addr +
-- 
2.31.1

[PATCH V9 14/19] virtio_ring: determine descriptor flags at one time

[Jason Wang]

Let's determine the last descriptor by counting the number of sg. This
would be consistent with packed virtqueue implementation and ease the
future in-order implementation.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 28 +++++++++++++---------------
 1 file changed, 13 insertions(+), 15 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index 4e905dc7b6cb..dbcd6f26df7d 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -575,7 +575,7 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
 	struct vring_desc_extra *extra;
 	struct scatterlist *sg;
 	struct vring_desc *desc;
-	unsigned int i, n, avail, descs_used, prev, err_idx;
+	unsigned int i, n, avail, descs_used, err_idx, sg_count = 0;
 	int head;
 	bool indirect;
 
@@ -635,42 +635,40 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
 		for (sg = sgs[n]; sg; sg = sg_next(sg)) {
 			dma_addr_t addr;
 			u32 len;
+			u16 flags = 0;
+
+			if (++sg_count != total_sg)
+				flags |= VRING_DESC_F_NEXT;
 
 			if (vring_map_one_sg(vq, sg, DMA_TO_DEVICE, &addr, &len, premapped))
 				goto unmap_release;
 
-			prev = i;
 			/* Note that we trust indirect descriptor
 			 * table since it use stream DMA mapping.
 			 */
-			i = virtqueue_add_desc_split(vq, desc, extra, i, addr, len,
-						     VRING_DESC_F_NEXT,
-						     premapped);
+			i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
+						     len, flags ,premapped);
 		}
 	}
 	for (; n < (out_sgs + in_sgs); n++) {
 		for (sg = sgs[n]; sg; sg = sg_next(sg)) {
 			dma_addr_t addr;
 			u32 len;
+			u16 flags = VRING_DESC_F_WRITE;
+
+			if (++sg_count != total_sg)
+				flags |= VRING_DESC_F_NEXT;
 
 			if (vring_map_one_sg(vq, sg, DMA_FROM_DEVICE, &addr, &len, premapped))
 				goto unmap_release;
 
-			prev = i;
 			/* Note that we trust indirect descriptor
 			 * table since it use stream DMA mapping.
 			 */
-			i = virtqueue_add_desc_split(vq, desc, extra, i, addr, len,
-						     VRING_DESC_F_NEXT |
-						     VRING_DESC_F_WRITE,
-						     premapped);
+			i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
+						     len, flags, premapped);
 		}
 	}
-	/* Last one doesn't continue. */
-	desc[prev].flags &= cpu_to_virtio16(vq->vq.vdev, ~VRING_DESC_F_NEXT);
-	if (!indirect && vring_need_unmap_buffer(vq, &extra[prev]))
-		vq->split.desc_extra[prev & (vq->split.vring.num - 1)].flags &=
-			~VRING_DESC_F_NEXT;
 
 	if (indirect) {
 		/* Now that the indirect table is filled in, map it. */
-- 
2.31.1

[PATCH V9 15/19] virtio_ring: factor out core logic of buffer detaching

[Jason Wang]

Factor out core logic of buffer detaching and leave the free list
management to the caller so in_order can just call the core logic.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index dbcd6f26df7d..ca33af830f65 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -1667,8 +1667,8 @@ static bool virtqueue_kick_prepare_packed(struct vring_virtqueue *vq)
 	return needs_kick;
 }
 
-static void detach_buf_packed(struct vring_virtqueue *vq,
-			      unsigned int id, void **ctx)
+static void detach_buf_packed_in_order(struct vring_virtqueue *vq,
+				       unsigned int id, void **ctx)
 {
 	struct vring_desc_state_packed *state = NULL;
 	struct vring_packed_desc *desc;
@@ -1679,8 +1679,6 @@ static void detach_buf_packed(struct vring_virtqueue *vq,
 	/* Clear data ptr. */
 	state->data = NULL;
 
-	vq->packed.desc_extra[state->last].next = vq->free_head;
-	vq->free_head = id;
 	vq->vq.num_free += state->num;
 
 	if (unlikely(vq->use_map_api)) {
@@ -1717,6 +1715,17 @@ static void detach_buf_packed(struct vring_virtqueue *vq,
 	}
 }
 
+static void detach_buf_packed(struct vring_virtqueue *vq,
+			      unsigned int id, void **ctx)
+{
+	struct vring_desc_state_packed *state = &vq->packed.desc_state[id];
+
+	vq->packed.desc_extra[state->last].next = vq->free_head;
+	vq->free_head = id;
+
+	detach_buf_packed_in_order(vq, id, ctx);
+}
+
 static inline bool is_used_desc_packed(const struct vring_virtqueue *vq,
 				       u16 idx, bool used_wrap_counter)
 {
-- 
2.31.1

[PATCH V9 16/19] virtio_ring: factor out core logic for updating last_used_idx

[Jason Wang]

Factor out the core logic for updating last_used_idx to be reused by
the packed in order implementation.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 43 +++++++++++++++++++++---------------
 1 file changed, 25 insertions(+), 18 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index ca33af830f65..9e3213f6083d 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -1756,6 +1756,30 @@ static bool more_used_packed(const struct vring_virtqueue *vq)
 	return virtqueue_poll_packed(vq, READ_ONCE(vq->last_used_idx));
 }
 
+static void update_last_used_idx_packed(struct vring_virtqueue *vq,
+					u16 id, u16 last_used,
+					u16 used_wrap_counter)
+{
+	last_used += vq->packed.desc_state[id].num;
+	if (unlikely(last_used >= vq->packed.vring.num)) {
+		last_used -= vq->packed.vring.num;
+		used_wrap_counter ^= 1;
+	}
+
+	last_used = (last_used | (used_wrap_counter << VRING_PACKED_EVENT_F_WRAP_CTR));
+	WRITE_ONCE(vq->last_used_idx, last_used);
+
+	/*
+	 * If we expect an interrupt for the next entry, tell host
+	 * by writing event index and flush out the write before
+	 * the read in the next get_buf call.
+	 */
+	if (vq->packed.event_flags_shadow == VRING_PACKED_EVENT_FLAG_DESC)
+		virtio_store_mb(vq->weak_barriers,
+				&vq->packed.vring.driver->off_wrap,
+				cpu_to_le16(vq->last_used_idx));
+}
+
 static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
 					  unsigned int *len,
 					  void **ctx)
@@ -1799,24 +1823,7 @@ static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
 	ret = vq->packed.desc_state[id].data;
 	detach_buf_packed(vq, id, ctx);
 
-	last_used += vq->packed.desc_state[id].num;
-	if (unlikely(last_used >= vq->packed.vring.num)) {
-		last_used -= vq->packed.vring.num;
-		used_wrap_counter ^= 1;
-	}
-
-	last_used = (last_used | (used_wrap_counter << VRING_PACKED_EVENT_F_WRAP_CTR));
-	WRITE_ONCE(vq->last_used_idx, last_used);
-
-	/*
-	 * If we expect an interrupt for the next entry, tell host
-	 * by writing event index and flush out the write before
-	 * the read in the next get_buf call.
-	 */
-	if (vq->packed.event_flags_shadow == VRING_PACKED_EVENT_FLAG_DESC)
-		virtio_store_mb(vq->weak_barriers,
-				&vq->packed.vring.driver->off_wrap,
-				cpu_to_le16(vq->last_used_idx));
+	update_last_used_idx_packed(vq, id, last_used, used_wrap_counter);
 
 	LAST_ADD_TIME_INVALID(vq);
 
-- 
2.31.1

[PATCH V9 17/19] virtio_ring: factor out split indirect detaching logic

[Jason Wang]

Factor out the split indirect descriptor detaching logic in order to
allow it to be reused by the in order support.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 62 ++++++++++++++++++++----------------
 1 file changed, 34 insertions(+), 28 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index 9e3213f6083d..5bb9797c6524 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -776,11 +776,41 @@ static bool virtqueue_kick_prepare_split(struct vring_virtqueue *vq)
 	return needs_kick;
 }
 
+static void detach_indirect_split(struct vring_virtqueue *vq,
+				  unsigned int head)
+{
+	struct vring_desc_extra *extra = vq->split.desc_extra;
+	struct vring_desc *indir_desc = vq->split.desc_state[head].indir_desc;
+	unsigned int j;
+	u32 len, num;
+
+	/* Free the indirect table, if any, now that it's unmapped. */
+	if (!indir_desc)
+		return;
+	len = vq->split.desc_extra[head].len;
+
+	BUG_ON(!(vq->split.desc_extra[head].flags &
+			VRING_DESC_F_INDIRECT));
+	BUG_ON(len == 0 || len % sizeof(struct vring_desc));
+
+	num = len / sizeof(struct vring_desc);
+
+	extra = (struct vring_desc_extra *)&indir_desc[num];
+
+	if (vq->use_map_api) {
+		for (j = 0; j < num; j++)
+			vring_unmap_one_split(vq, &extra[j]);
+	}
+
+	kfree(indir_desc);
+	vq->split.desc_state[head].indir_desc = NULL;
+}
+
 static void detach_buf_split(struct vring_virtqueue *vq, unsigned int head,
 			     void **ctx)
 {
 	struct vring_desc_extra *extra;
-	unsigned int i, j;
+	unsigned int i;
 	__virtio16 nextflag = cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_NEXT);
 
 	/* Clear data ptr. */
@@ -804,34 +834,10 @@ static void detach_buf_split(struct vring_virtqueue *vq, unsigned int head,
 	/* Plus final descriptor */
 	vq->vq.num_free++;
 
-	if (vq->indirect) {
-		struct vring_desc *indir_desc =
-				vq->split.desc_state[head].indir_desc;
-		u32 len, num;
-
-		/* Free the indirect table, if any, now that it's unmapped. */
-		if (!indir_desc)
-			return;
-		len = vq->split.desc_extra[head].len;
-
-		BUG_ON(!(vq->split.desc_extra[head].flags &
-				VRING_DESC_F_INDIRECT));
-		BUG_ON(len == 0 || len % sizeof(struct vring_desc));
-
-		num = len / sizeof(struct vring_desc);
-
-		extra = (struct vring_desc_extra *)&indir_desc[num];
-
-		if (vq->use_map_api) {
-			for (j = 0; j < num; j++)
-				vring_unmap_one_split(vq, &extra[j]);
-		}
-
-		kfree(indir_desc);
-		vq->split.desc_state[head].indir_desc = NULL;
-	} else if (ctx) {
+	if (vq->indirect)
+		detach_indirect_split(vq, head);
+	else if (ctx)
 		*ctx = vq->split.desc_state[head].indir_desc;
-	}
 }
 
 static bool virtqueue_poll_split(const struct vring_virtqueue *vq,
-- 
2.31.1

[PATCH V9 18/19] virtio_ring: factor out split detaching logic

[Jason Wang]

This patch factors out the split core detaching logic that could be
reused by in order feature into a dedicated function.

Acked-by: Eugenio Pérez <eperezma@redhat.com>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index 5bb9797c6524..61884e031b94 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -806,8 +806,9 @@ static void detach_indirect_split(struct vring_virtqueue *vq,
 	vq->split.desc_state[head].indir_desc = NULL;
 }
 
-static void detach_buf_split(struct vring_virtqueue *vq, unsigned int head,
-			     void **ctx)
+static unsigned detach_buf_split_in_order(struct vring_virtqueue *vq,
+					  unsigned int head,
+					  void **ctx)
 {
 	struct vring_desc_extra *extra;
 	unsigned int i;
@@ -828,8 +829,6 @@ static void detach_buf_split(struct vring_virtqueue *vq, unsigned int head,
 	}
 
 	vring_unmap_one_split(vq, &extra[i]);
-	vq->split.desc_extra[i].next = vq->free_head;
-	vq->free_head = head;
 
 	/* Plus final descriptor */
 	vq->vq.num_free++;
@@ -838,6 +837,17 @@ static void detach_buf_split(struct vring_virtqueue *vq, unsigned int head,
 		detach_indirect_split(vq, head);
 	else if (ctx)
 		*ctx = vq->split.desc_state[head].indir_desc;
+
+	return i;
+}
+
+static void detach_buf_split(struct vring_virtqueue *vq, unsigned int head,
+			     void **ctx)
+{
+	unsigned int i = detach_buf_split_in_order(vq, head, ctx);
+
+	vq->split.desc_extra[i].next = vq->free_head;
+	vq->free_head = head;
 }
 
 static bool virtqueue_poll_split(const struct vring_virtqueue *vq,
-- 
2.31.1

[PATCH V9 19/19] virtio_ring: add in order support

[Jason Wang]

This patch implements in order support for both split virtqueue and
packed virtqueue. Performance could be gained for the device where the
memory access could be expensive (e.g vhost-net or a real PCI device):

Benchmark with KVM guest:

Vhost-net on the host: (pktgen + XDP_DROP):

         in_order=off | in_order=on | +%
    TX:  4.51Mpps     | 5.30Mpps    | +17%
    RX:  3.47Mpps     | 3.61Mpps    | + 4%

Vhost-user(testpmd) on the host: (pktgen/XDP_DROP):

For split virtqueue:

         in_order=off | in_order=on | +%
    TX:  5.60Mpps     | 5.60Mpps    | +0.0%
    RX:  9.16Mpps     | 9.61Mpps    | +4.9%

For packed virtqueue:

         in_order=off | in_order=on | +%
    TX:  5.60Mpps     | 5.70Mpps    | +1.7%
    RX:  10.6Mpps     | 10.8Mpps    | +1.8%

Benchmark also shows no performance impact for in_order=off for queue
size with 256 and 1024.

Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/virtio/virtio_ring.c | 458 +++++++++++++++++++++++++++++++++--
 1 file changed, 435 insertions(+), 23 deletions(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index 61884e031b94..9847b23f82ea 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -70,6 +70,8 @@
 enum vq_layout {
 	SPLIT = 0,
 	PACKED,
+	SPLIT_IN_ORDER,
+	PACKED_IN_ORDER,
 	VQ_TYPE_MAX,
 };
 
@@ -80,6 +82,7 @@ struct vring_desc_state_split {
 	 * allocated together. So we won't stress more to the memory allocator.
 	 */
 	struct vring_desc *indir_desc;
+	u32 total_len;			/* Buffer Length */
 };
 
 struct vring_desc_state_packed {
@@ -91,6 +94,7 @@ struct vring_desc_state_packed {
 	struct vring_packed_desc *indir_desc;
 	u16 num;			/* Descriptor list length. */
 	u16 last;			/* The last desc state in a list. */
+	u32 total_len;			/* Buffer Length */
 };
 
 struct vring_desc_extra {
@@ -205,8 +209,24 @@ struct vring_virtqueue {
 
 	enum vq_layout layout;
 
-	/* Head of free buffer list. */
+	/*
+	 * Without IN_ORDER it's the head of free buffer list. With
+	 * IN_ORDER and SPLIT, it's the next available buffer
+	 * index. With IN_ORDER and PACKED, it's unused.
+	 */
 	unsigned int free_head;
+
+	/*
+	 * With IN_ORDER, once we see an in-order batch, this stores
+	 * this last entry, and until we return the last buffer.
+	 * After this, id is set to vq.num to mark it invalid.
+	 * Unused without IN_ORDER.
+	 */
+	struct used_entry {
+		u32 id;
+		u32 len;
+	} batch_last;
+
 	/* Number we've added since last sync. */
 	unsigned int num_added;
 
@@ -218,6 +238,11 @@ struct vring_virtqueue {
 	 */
 	u16 last_used_idx;
 
+	/* With IN_ORDER and SPLIT, last descriptor id we used to
+	 * detach buffer.
+	 */
+	u16 last_used;
+
 	/* Hint for event idx: already triggered no need to disable. */
 	bool event_triggered;
 
@@ -259,7 +284,12 @@ static void vring_free(struct virtqueue *_vq);
 
 static inline bool virtqueue_is_packed(const struct vring_virtqueue *vq)
 {
-	return vq->layout == PACKED;
+	return vq->layout == PACKED || vq->layout == PACKED_IN_ORDER;
+}
+
+static inline bool virtqueue_is_in_order(const struct vring_virtqueue *vq)
+{
+	return vq->layout == SPLIT_IN_ORDER || vq->layout == PACKED_IN_ORDER;
 }
 
 static bool virtqueue_use_indirect(const struct vring_virtqueue *vq,
@@ -469,6 +499,8 @@ static void virtqueue_init(struct vring_virtqueue *vq, u32 num)
 	else
 		vq->last_used_idx = 0;
 
+	vq->last_used = 0;
+
 	vq->event_triggered = false;
 	vq->num_added = 0;
 
@@ -576,6 +608,8 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
 	struct scatterlist *sg;
 	struct vring_desc *desc;
 	unsigned int i, n, avail, descs_used, err_idx, sg_count = 0;
+	/* Total length for in-order */
+	unsigned int total_len = 0;
 	int head;
 	bool indirect;
 
@@ -648,6 +682,7 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
 			 */
 			i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
 						     len, flags ,premapped);
+			total_len += len;
 		}
 	}
 	for (; n < (out_sgs + in_sgs); n++) {
@@ -667,6 +702,7 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
 			 */
 			i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
 						     len, flags, premapped);
+			total_len += len;
 		}
 	}
 
@@ -689,7 +725,12 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
 	vq->vq.num_free -= descs_used;
 
 	/* Update free pointer */
-	if (indirect)
+	if (virtqueue_is_in_order(vq)) {
+		vq->free_head += descs_used;
+		if (vq->free_head >= vq->split.vring.num)
+			vq->free_head -= vq->split.vring.num;
+		vq->split.desc_state[head].total_len = total_len;
+	} else if (indirect)
 		vq->free_head = vq->split.desc_extra[head].next;
 	else
 		vq->free_head = i;
@@ -862,6 +903,14 @@ static bool more_used_split(const struct vring_virtqueue *vq)
 	return virtqueue_poll_split(vq, vq->last_used_idx);
 }
 
+static bool more_used_split_in_order(const struct vring_virtqueue *vq)
+{
+	if (vq->batch_last.id != UINT_MAX)
+		return true;
+
+	return virtqueue_poll_split(vq, vq->last_used_idx);
+}
+
 static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
 					 unsigned int *len,
 					 void **ctx)
@@ -919,6 +968,80 @@ static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
 	return ret;
 }
 
+static void *virtqueue_get_buf_ctx_split_in_order(struct vring_virtqueue *vq,
+						  unsigned int *len,
+						  void **ctx)
+{
+	void *ret;
+	unsigned int num = vq->split.vring.num;
+	unsigned int num_free = vq->vq.num_free;
+	u16 last_used, last_used_idx;
+
+	START_USE(vq);
+
+	if (unlikely(vq->broken)) {
+		END_USE(vq);
+		return NULL;
+	}
+
+	last_used = vq->last_used & (num - 1);
+	last_used_idx = vq->last_used_idx & (num - 1);
+
+	if (vq->batch_last.id == UINT_MAX) {
+		if (!more_used_split_in_order(vq)) {
+			pr_debug("No more buffers in queue\n");
+			END_USE(vq);
+			return NULL;
+		}
+
+		/*
+		 * Only get used array entries after they have been
+		 * exposed by host.
+		 */
+		virtio_rmb(vq->weak_barriers);
+
+		vq->batch_last.id = virtio32_to_cpu(vq->vq.vdev,
+				    vq->split.vring.used->ring[last_used_idx].id);
+		vq->batch_last.len = virtio32_to_cpu(vq->vq.vdev,
+				     vq->split.vring.used->ring[last_used_idx].len);
+	}
+
+	if (vq->batch_last.id == last_used) {
+		vq->batch_last.id = UINT_MAX;
+		*len = vq->batch_last.len;
+	} else {
+		*len = vq->split.desc_state[last_used].total_len;
+	}
+
+	if (unlikely(last_used >= num)) {
+		BAD_RING(vq, "id %u out of range\n", last_used);
+		return NULL;
+	}
+	if (unlikely(!vq->split.desc_state[last_used].data)) {
+		BAD_RING(vq, "id %u is not a head!\n", last_used);
+		return NULL;
+	}
+
+	/* detach_buf_split clears data, so grab it now. */
+	ret = vq->split.desc_state[last_used].data;
+	detach_buf_split_in_order(vq, last_used, ctx);
+
+	vq->last_used_idx++;
+	vq->last_used += (vq->vq.num_free - num_free);
+	/* If we expect an interrupt for the next entry, tell host
+	 * by writing event index and flush out the write before
+	 * the read in the next get_buf call. */
+	if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
+		virtio_store_mb(vq->weak_barriers,
+				&vring_used_event(&vq->split.vring),
+				cpu_to_virtio16(vq->vq.vdev, vq->last_used_idx));
+
+	LAST_ADD_TIME_INVALID(vq);
+
+	END_USE(vq);
+	return ret;
+}
+
 static void virtqueue_disable_cb_split(struct vring_virtqueue *vq)
 {
 	if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
@@ -1012,7 +1135,10 @@ static void *virtqueue_detach_unused_buf_split(struct vring_virtqueue *vq)
 			continue;
 		/* detach_buf_split clears data, so grab it now. */
 		buf = vq->split.desc_state[i].data;
-		detach_buf_split(vq, i, NULL);
+		if (virtqueue_is_in_order(vq))
+			detach_buf_split_in_order(vq, i, NULL);
+		else
+			detach_buf_split(vq, i, NULL);
 		vq->split.avail_idx_shadow--;
 		vq->split.vring.avail->idx = cpu_to_virtio16(vq->vq.vdev,
 				vq->split.avail_idx_shadow);
@@ -1075,6 +1201,7 @@ static void virtqueue_vring_attach_split(struct vring_virtqueue *vq,
 
 	/* Put everything in free lists. */
 	vq->free_head = 0;
+	vq->batch_last.id = UINT_MAX;
 }
 
 static int vring_alloc_state_extra_split(struct vring_virtqueue_split *vring_split)
@@ -1186,7 +1313,6 @@ static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
 	if (!vq)
 		return NULL;
 
-	vq->layout = SPLIT;
 	vq->vq.callback = callback;
 	vq->vq.vdev = vdev;
 	vq->vq.name = name;
@@ -1206,6 +1332,8 @@ static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
 	vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
 		!context;
 	vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
+	vq->layout = virtio_has_feature(vdev, VIRTIO_F_IN_ORDER) ?
+		     SPLIT_IN_ORDER : SPLIT;
 
 	if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
 		vq->weak_barriers = false;
@@ -1363,13 +1491,14 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
 					 unsigned int in_sgs,
 					 void *data,
 					 bool premapped,
-					 gfp_t gfp)
+					 gfp_t gfp,
+					 u16 id)
 {
 	struct vring_desc_extra *extra;
 	struct vring_packed_desc *desc;
 	struct scatterlist *sg;
-	unsigned int i, n, err_idx, len;
-	u16 head, id;
+	unsigned int i, n, err_idx, len, total_len = 0;
+	u16 head;
 	dma_addr_t addr;
 
 	head = vq->packed.next_avail_idx;
@@ -1387,8 +1516,6 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
 	}
 
 	i = 0;
-	id = vq->free_head;
-	BUG_ON(id == vq->packed.vring.num);
 
 	for (n = 0; n < out_sgs + in_sgs; n++) {
 		for (sg = sgs[n]; sg; sg = sg_next(sg)) {
@@ -1408,6 +1535,7 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
 				extra[i].flags = n < out_sgs ?  0 : VRING_DESC_F_WRITE;
 			}
 
+			total_len += len;
 			i++;
 		}
 	}
@@ -1454,13 +1582,15 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
 				1 << VRING_PACKED_DESC_F_USED;
 	}
 	vq->packed.next_avail_idx = n;
-	vq->free_head = vq->packed.desc_extra[id].next;
+	if (!virtqueue_is_in_order(vq))
+		vq->free_head = vq->packed.desc_extra[id].next;
 
 	/* Store token and indirect buffer state. */
 	vq->packed.desc_state[id].num = 1;
 	vq->packed.desc_state[id].data = data;
 	vq->packed.desc_state[id].indir_desc = desc;
 	vq->packed.desc_state[id].last = id;
+	vq->packed.desc_state[id].total_len = total_len;
 
 	vq->num_added += 1;
 
@@ -1513,8 +1643,11 @@ static inline int virtqueue_add_packed(struct vring_virtqueue *vq,
 	BUG_ON(total_sg == 0);
 
 	if (virtqueue_use_indirect(vq, total_sg)) {
+		id = vq->free_head;
+		BUG_ON(id == vq->packed.vring.num);
 		err = virtqueue_add_indirect_packed(vq, sgs, total_sg, out_sgs,
-						    in_sgs, data, premapped, gfp);
+						    in_sgs, data, premapped,
+						    gfp, id);
 		if (err != -ENOMEM) {
 			END_USE(vq);
 			return err;
@@ -1635,6 +1768,156 @@ static inline int virtqueue_add_packed(struct vring_virtqueue *vq,
 	return -EIO;
 }
 
+static inline int virtqueue_add_packed_in_order(struct vring_virtqueue *vq,
+						struct scatterlist *sgs[],
+						unsigned int total_sg,
+						unsigned int out_sgs,
+						unsigned int in_sgs,
+						void *data,
+						void *ctx,
+						bool premapped,
+						gfp_t gfp)
+{
+	struct vring_packed_desc *desc;
+	struct scatterlist *sg;
+	unsigned int i, n, sg_count, err_idx, total_len = 0;
+	__le16 head_flags, flags;
+	u16 head, avail_used_flags;
+	int err;
+
+	START_USE(vq);
+
+	BUG_ON(data == NULL);
+	BUG_ON(ctx && vq->indirect);
+
+	if (unlikely(vq->broken)) {
+		END_USE(vq);
+		return -EIO;
+	}
+
+	LAST_ADD_TIME_UPDATE(vq);
+
+	BUG_ON(total_sg == 0);
+
+	if (virtqueue_use_indirect(vq, total_sg)) {
+		err = virtqueue_add_indirect_packed(vq, sgs, total_sg, out_sgs,
+						    in_sgs, data, premapped, gfp,
+						    vq->packed.next_avail_idx);
+		if (err != -ENOMEM) {
+			END_USE(vq);
+			return err;
+		}
+
+		/* fall back on direct */
+	}
+
+	head = vq->packed.next_avail_idx;
+	avail_used_flags = vq->packed.avail_used_flags;
+
+	WARN_ON_ONCE(total_sg > vq->packed.vring.num && !vq->indirect);
+
+	desc = vq->packed.vring.desc;
+	i = head;
+
+	if (unlikely(vq->vq.num_free < total_sg)) {
+		pr_debug("Can't add buf len %i - avail = %i\n",
+			 total_sg, vq->vq.num_free);
+		END_USE(vq);
+		return -ENOSPC;
+	}
+
+	sg_count = 0;
+	for (n = 0; n < out_sgs + in_sgs; n++) {
+		for (sg = sgs[n]; sg; sg = sg_next(sg)) {
+			dma_addr_t addr;
+			u32 len;
+
+			flags = 0;
+			if (++sg_count != total_sg)
+				flags |= cpu_to_le16(VRING_DESC_F_NEXT);
+			if (n >= out_sgs)
+				flags |= cpu_to_le16(VRING_DESC_F_WRITE);
+
+			if (vring_map_one_sg(vq, sg, n < out_sgs ?
+					     DMA_TO_DEVICE : DMA_FROM_DEVICE,
+					     &addr, &len, premapped))
+				goto unmap_release;
+
+			flags |= cpu_to_le16(vq->packed.avail_used_flags);
+
+			if (i == head)
+				head_flags = flags;
+			else
+				desc[i].flags = flags;
+
+			desc[i].addr = cpu_to_le64(addr);
+			desc[i].len = cpu_to_le32(len);
+			desc[i].id = cpu_to_le16(head);
+
+			if (unlikely(vq->use_map_api)) {
+				vq->packed.desc_extra[i].addr = premapped ?
+				      DMA_MAPPING_ERROR: addr;
+				vq->packed.desc_extra[i].len = len;
+				vq->packed.desc_extra[i].flags =
+					le16_to_cpu(flags);
+			}
+
+			if ((unlikely(++i >= vq->packed.vring.num))) {
+				i = 0;
+				vq->packed.avail_used_flags ^=
+					1 << VRING_PACKED_DESC_F_AVAIL |
+					1 << VRING_PACKED_DESC_F_USED;
+				vq->packed.avail_wrap_counter ^= 1;
+			}
+
+			total_len += len;
+		}
+	}
+
+	/* We're using some buffers from the free list. */
+	vq->vq.num_free -= total_sg;
+
+	/* Update free pointer */
+	vq->packed.next_avail_idx = i;
+
+	/* Store token. */
+	vq->packed.desc_state[head].num = total_sg;
+	vq->packed.desc_state[head].data = data;
+	vq->packed.desc_state[head].indir_desc = ctx;
+	vq->packed.desc_state[head].total_len = total_len;
+
+	/*
+	 * A driver MUST NOT make the first descriptor in the list
+	 * available before all subsequent descriptors comprising
+	 * the list are made available.
+	 */
+	virtio_wmb(vq->weak_barriers);
+	vq->packed.vring.desc[head].flags = head_flags;
+	vq->num_added += total_sg;
+
+	pr_debug("Added buffer head %i to %p\n", head, vq);
+	END_USE(vq);
+
+	return 0;
+
+unmap_release:
+	err_idx = i;
+	i = head;
+	vq->packed.avail_used_flags = avail_used_flags;
+
+	for (n = 0; n < total_sg; n++) {
+		if (i == err_idx)
+			break;
+		vring_unmap_extra_packed(vq, &vq->packed.desc_extra[i]);
+		i++;
+		if (i >= vq->packed.vring.num)
+			i = 0;
+	}
+
+	END_USE(vq);
+	return -EIO;
+}
+
 static bool virtqueue_kick_prepare_packed(struct vring_virtqueue *vq)
 {
 	u16 new, old, off_wrap, flags, wrap_counter, event_idx;
@@ -1796,10 +2079,82 @@ static void update_last_used_idx_packed(struct vring_virtqueue *vq,
 				cpu_to_le16(vq->last_used_idx));
 }
 
+static bool more_used_packed_in_order(const struct vring_virtqueue *vq)
+{
+	if (vq->batch_last.id != UINT_MAX)
+		return true;
+
+	return virtqueue_poll_packed(vq, READ_ONCE(vq->last_used_idx));
+}
+
+static void *virtqueue_get_buf_ctx_packed_in_order(struct vring_virtqueue *vq,
+						   unsigned int *len,
+						   void **ctx)
+{
+	unsigned int num = vq->packed.vring.num;
+	u16 last_used, last_used_idx;
+	bool used_wrap_counter;
+	void *ret;
+
+	START_USE(vq);
+
+	if (unlikely(vq->broken)) {
+		END_USE(vq);
+		return NULL;
+	}
+
+	last_used_idx = vq->last_used_idx;
+	used_wrap_counter = packed_used_wrap_counter(last_used_idx);
+	last_used = packed_last_used(last_used_idx);
+
+	if (vq->batch_last.id == UINT_MAX) {
+		if (!more_used_packed_in_order(vq)) {
+			pr_debug("No more buffers in queue\n");
+			END_USE(vq);
+			return NULL;
+		}
+		/* Only get used elements after they have been exposed by host. */
+		virtio_rmb(vq->weak_barriers);
+		vq->batch_last.id =
+			le16_to_cpu(vq->packed.vring.desc[last_used].id);
+		vq->batch_last.len =
+			le32_to_cpu(vq->packed.vring.desc[last_used].len);
+	}
+
+	if (vq->batch_last.id == last_used) {
+		vq->batch_last.id = UINT_MAX;
+		*len = vq->batch_last.len;
+	} else {
+		*len = vq->packed.desc_state[last_used].total_len;
+	}
+
+	if (unlikely(last_used >= num)) {
+		BAD_RING(vq, "id %u out of range\n", last_used);
+		return NULL;
+	}
+	if (unlikely(!vq->packed.desc_state[last_used].data)) {
+		BAD_RING(vq, "id %u is not a head!\n", last_used);
+		return NULL;
+	}
+
+	/* detach_buf_packed clears data, so grab it now. */
+	ret = vq->packed.desc_state[last_used].data;
+	detach_buf_packed_in_order(vq, last_used, ctx);
+
+	update_last_used_idx_packed(vq, last_used, last_used,
+				    used_wrap_counter);
+
+	LAST_ADD_TIME_INVALID(vq);
+
+	END_USE(vq);
+	return ret;
+}
+
 static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
 					  unsigned int *len,
 					  void **ctx)
 {
+	unsigned int num = vq->packed.vring.num;
 	u16 last_used, id, last_used_idx;
 	bool used_wrap_counter;
 	void *ret;
@@ -1826,7 +2181,7 @@ static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
 	id = le16_to_cpu(vq->packed.vring.desc[last_used].id);
 	*len = le32_to_cpu(vq->packed.vring.desc[last_used].len);
 
-	if (unlikely(id >= vq->packed.vring.num)) {
+	if (unlikely(id >= num)) {
 		BAD_RING(vq, "id %u out of range\n", id);
 		return NULL;
 	}
@@ -1967,7 +2322,10 @@ static void *virtqueue_detach_unused_buf_packed(struct vring_virtqueue *vq)
 			continue;
 		/* detach_buf clears data, so grab it now. */
 		buf = vq->packed.desc_state[i].data;
-		detach_buf_packed(vq, i, NULL);
+		if (virtqueue_is_in_order(vq))
+			detach_buf_packed_in_order(vq, i, NULL);
+		else
+			detach_buf_packed(vq, i, NULL);
 		END_USE(vq);
 		return buf;
 	}
@@ -1993,6 +2351,8 @@ static struct vring_desc_extra *vring_alloc_desc_extra(unsigned int num)
 	for (i = 0; i < num - 1; i++)
 		desc_extra[i].next = i + 1;
 
+	desc_extra[num - 1].next = 0;
+
 	return desc_extra;
 }
 
@@ -2124,10 +2484,17 @@ static void virtqueue_vring_attach_packed(struct vring_virtqueue *vq,
 {
 	vq->packed = *vring_packed;
 
-	/* Put everything in free lists. */
-	vq->free_head = 0;
+	if (virtqueue_is_in_order(vq)) {
+		vq->batch_last.id = UINT_MAX;
+	} else {
+		/*
+		 * Put everything in free lists. Note that
+		 * next_avail_idx is sufficient with IN_ORDER so
+		 * free_head is unused.
+		 */
+		vq->free_head = 0;
+	}
 }
-
 static void virtqueue_reset_packed(struct vring_virtqueue *vq)
 {
 	memset(vq->packed.vring.device, 0, vq->packed.event_size_in_bytes);
@@ -2172,13 +2539,14 @@ static struct virtqueue *__vring_new_virtqueue_packed(unsigned int index,
 #else
 	vq->broken = false;
 #endif
-	vq->layout = PACKED;
 	vq->map = map;
 	vq->use_map_api = vring_use_map_api(vdev);
 
 	vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
 		!context;
 	vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
+	vq->layout = virtio_has_feature(vdev, VIRTIO_F_IN_ORDER) ?
+		     PACKED_IN_ORDER : PACKED;
 
 	if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
 		vq->weak_barriers = false;
@@ -2288,9 +2656,39 @@ static const struct virtqueue_ops packed_ops = {
 	.reset = virtqueue_reset_packed,
 };
 
+static const struct virtqueue_ops split_in_order_ops = {
+	.add = virtqueue_add_split,
+	.get = virtqueue_get_buf_ctx_split_in_order,
+	.kick_prepare = virtqueue_kick_prepare_split,
+	.disable_cb = virtqueue_disable_cb_split,
+	.enable_cb_delayed = virtqueue_enable_cb_delayed_split,
+	.enable_cb_prepare = virtqueue_enable_cb_prepare_split,
+	.poll = virtqueue_poll_split,
+	.detach_unused_buf = virtqueue_detach_unused_buf_split,
+	.more_used = more_used_split_in_order,
+	.resize = virtqueue_resize_split,
+	.reset = virtqueue_reset_split,
+};
+
+static const struct virtqueue_ops packed_in_order_ops = {
+	.add = virtqueue_add_packed_in_order,
+	.get = virtqueue_get_buf_ctx_packed_in_order,
+	.kick_prepare = virtqueue_kick_prepare_packed,
+	.disable_cb = virtqueue_disable_cb_packed,
+	.enable_cb_delayed = virtqueue_enable_cb_delayed_packed,
+	.enable_cb_prepare = virtqueue_enable_cb_prepare_packed,
+	.poll = virtqueue_poll_packed,
+	.detach_unused_buf = virtqueue_detach_unused_buf_packed,
+	.more_used = more_used_packed_in_order,
+	.resize = virtqueue_resize_packed,
+	.reset = virtqueue_reset_packed,
+};
+
 static const struct virtqueue_ops *const all_ops[VQ_TYPE_MAX] = {
 	[SPLIT] = &split_ops,
-	[PACKED] = &packed_ops
+	[PACKED] = &packed_ops,
+	[SPLIT_IN_ORDER] = &split_in_order_ops,
+	[PACKED_IN_ORDER] = &packed_in_order_ops,
 };
 
 static int virtqueue_disable_and_recycle(struct virtqueue *_vq,
@@ -2346,6 +2744,12 @@ static int virtqueue_enable_after_reset(struct virtqueue *_vq)
 	case PACKED:							\
 		ret = all_ops[PACKED]->op(vq, ##__VA_ARGS__);		\
 		break;							\
+	case SPLIT_IN_ORDER:						\
+		ret = all_ops[SPLIT_IN_ORDER]->op(vq, ##__VA_ARGS__);	\
+		break;							\
+	case PACKED_IN_ORDER:						\
+		ret = all_ops[PACKED_IN_ORDER]->op(vq, ##__VA_ARGS__);	\
+		break;							\
 	default:							\
 		BUG();							\
 		break;							\
@@ -2362,10 +2766,16 @@ static int virtqueue_enable_after_reset(struct virtqueue *_vq)
 	case PACKED:					\
 		all_ops[PACKED]->op(vq, ##__VA_ARGS__);	\
 		break;					\
-	default:					\
-		BUG();					\
-		break;					\
-	}						\
+	case SPLIT_IN_ORDER:						\
+		all_ops[SPLIT_IN_ORDER]->op(vq, ##__VA_ARGS__);	\
+		break;							\
+	case PACKED_IN_ORDER:						\
+		all_ops[PACKED_IN_ORDER]->op(vq, ##__VA_ARGS__);	\
+		break;							\
+	default:							\
+		BUG();							\
+		break;							\
+	}								\
 })
 
 static inline int virtqueue_add(struct virtqueue *_vq,
@@ -3081,6 +3491,8 @@ void vring_transport_features(struct virtio_device *vdev)
 			break;
 		case VIRTIO_F_NOTIFICATION_DATA:
 			break;
+		case VIRTIO_F_IN_ORDER:
+			break;
 		default:
 			/* We don't understand this bit. */
 			__virtio_clear_bit(vdev, i);
-- 
2.31.1

Re: [PATCH V9 19/19] virtio_ring: add in order support

[Eugenio Perez Martin]

On Tue, Nov 25, 2025 at 8:10 AM Jason Wang <jasowang@redhat.com> wrote:
>
> This patch implements in order support for both split virtqueue and
> packed virtqueue. Performance could be gained for the device where the
> memory access could be expensive (e.g vhost-net or a real PCI device):
>
> Benchmark with KVM guest:
>
> Vhost-net on the host: (pktgen + XDP_DROP):
>
>          in_order=off | in_order=on | +%
>     TX:  4.51Mpps     | 5.30Mpps    | +17%
>     RX:  3.47Mpps     | 3.61Mpps    | + 4%
>
> Vhost-user(testpmd) on the host: (pktgen/XDP_DROP):
>
> For split virtqueue:
>
>          in_order=off | in_order=on | +%
>     TX:  5.60Mpps     | 5.60Mpps    | +0.0%
>     RX:  9.16Mpps     | 9.61Mpps    | +4.9%
>
> For packed virtqueue:
>
>          in_order=off | in_order=on | +%
>     TX:  5.60Mpps     | 5.70Mpps    | +1.7%
>     RX:  10.6Mpps     | 10.8Mpps    | +1.8%
>
> Benchmark also shows no performance impact for in_order=off for queue
> size with 256 and 1024.
>
> Signed-off-by: Jason Wang <jasowang@redhat.com>

Reviewed-by: Eugenio Pérez <eperezma@redhat.com>

I have one nitpick though, but it can be added on top or just ignored.

> ---
>  drivers/virtio/virtio_ring.c | 458 +++++++++++++++++++++++++++++++++--
>  1 file changed, 435 insertions(+), 23 deletions(-)
>
> diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> index 61884e031b94..9847b23f82ea 100644
> --- a/drivers/virtio/virtio_ring.c
> +++ b/drivers/virtio/virtio_ring.c
> @@ -70,6 +70,8 @@
>  enum vq_layout {
>         SPLIT = 0,
>         PACKED,
> +       SPLIT_IN_ORDER,
> +       PACKED_IN_ORDER,
>         VQ_TYPE_MAX,
>  };
>
> @@ -80,6 +82,7 @@ struct vring_desc_state_split {
>          * allocated together. So we won't stress more to the memory allocator.
>          */
>         struct vring_desc *indir_desc;
> +       u32 total_len;                  /* Buffer Length */
>  };
>
>  struct vring_desc_state_packed {
> @@ -91,6 +94,7 @@ struct vring_desc_state_packed {
>         struct vring_packed_desc *indir_desc;
>         u16 num;                        /* Descriptor list length. */
>         u16 last;                       /* The last desc state in a list. */
> +       u32 total_len;                  /* Buffer Length */
>  };
>
>  struct vring_desc_extra {
> @@ -205,8 +209,24 @@ struct vring_virtqueue {
>
>         enum vq_layout layout;
>
> -       /* Head of free buffer list. */
> +       /*
> +        * Without IN_ORDER it's the head of free buffer list. With
> +        * IN_ORDER and SPLIT, it's the next available buffer
> +        * index. With IN_ORDER and PACKED, it's unused.
> +        */
>         unsigned int free_head;
> +
> +       /*
> +        * With IN_ORDER, once we see an in-order batch, this stores
> +        * this last entry, and until we return the last buffer.
> +        * After this, id is set to vq.num to mark it invalid.
> +        * Unused without IN_ORDER.
> +        */
> +       struct used_entry {
> +               u32 id;
> +               u32 len;
> +       } batch_last;
> +
>         /* Number we've added since last sync. */
>         unsigned int num_added;
>
> @@ -218,6 +238,11 @@ struct vring_virtqueue {
>          */
>         u16 last_used_idx;
>
> +       /* With IN_ORDER and SPLIT, last descriptor id we used to
> +        * detach buffer.
> +        */
> +       u16 last_used;
> +
>         /* Hint for event idx: already triggered no need to disable. */
>         bool event_triggered;
>
> @@ -259,7 +284,12 @@ static void vring_free(struct virtqueue *_vq);
>
>  static inline bool virtqueue_is_packed(const struct vring_virtqueue *vq)
>  {
> -       return vq->layout == PACKED;
> +       return vq->layout == PACKED || vq->layout == PACKED_IN_ORDER;
> +}
> +
> +static inline bool virtqueue_is_in_order(const struct vring_virtqueue *vq)
> +{
> +       return vq->layout == SPLIT_IN_ORDER || vq->layout == PACKED_IN_ORDER;
>  }
>
>  static bool virtqueue_use_indirect(const struct vring_virtqueue *vq,
> @@ -469,6 +499,8 @@ static void virtqueue_init(struct vring_virtqueue *vq, u32 num)
>         else
>                 vq->last_used_idx = 0;
>
> +       vq->last_used = 0;
> +
>         vq->event_triggered = false;
>         vq->num_added = 0;
>
> @@ -576,6 +608,8 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
>         struct scatterlist *sg;
>         struct vring_desc *desc;
>         unsigned int i, n, avail, descs_used, err_idx, sg_count = 0;
> +       /* Total length for in-order */
> +       unsigned int total_len = 0;
>         int head;
>         bool indirect;
>
> @@ -648,6 +682,7 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
>                          */
>                         i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
>                                                      len, flags ,premapped);
> +                       total_len += len;
>                 }
>         }
>         for (; n < (out_sgs + in_sgs); n++) {
> @@ -667,6 +702,7 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
>                          */
>                         i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
>                                                      len, flags, premapped);
> +                       total_len += len;
>                 }
>         }
>
> @@ -689,7 +725,12 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
>         vq->vq.num_free -= descs_used;
>
>         /* Update free pointer */
> -       if (indirect)
> +       if (virtqueue_is_in_order(vq)) {
> +               vq->free_head += descs_used;
> +               if (vq->free_head >= vq->split.vring.num)
> +                       vq->free_head -= vq->split.vring.num;
> +               vq->split.desc_state[head].total_len = total_len;
> +       } else if (indirect)
>                 vq->free_head = vq->split.desc_extra[head].next;
>         else
>                 vq->free_head = i;
> @@ -862,6 +903,14 @@ static bool more_used_split(const struct vring_virtqueue *vq)
>         return virtqueue_poll_split(vq, vq->last_used_idx);
>  }
>
> +static bool more_used_split_in_order(const struct vring_virtqueue *vq)
> +{
> +       if (vq->batch_last.id != UINT_MAX)
> +               return true;
> +
> +       return virtqueue_poll_split(vq, vq->last_used_idx);
> +}
> +
>  static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
>                                          unsigned int *len,
>                                          void **ctx)
> @@ -919,6 +968,80 @@ static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
>         return ret;
>  }
>
> +static void *virtqueue_get_buf_ctx_split_in_order(struct vring_virtqueue *vq,
> +                                                 unsigned int *len,
> +                                                 void **ctx)
> +{
> +       void *ret;
> +       unsigned int num = vq->split.vring.num;
> +       unsigned int num_free = vq->vq.num_free;
> +       u16 last_used, last_used_idx;
> +
> +       START_USE(vq);
> +
> +       if (unlikely(vq->broken)) {
> +               END_USE(vq);
> +               return NULL;
> +       }
> +
> +       last_used = vq->last_used & (num - 1);
> +       last_used_idx = vq->last_used_idx & (num - 1);
> +
> +       if (vq->batch_last.id == UINT_MAX) {
> +               if (!more_used_split_in_order(vq)) {
> +                       pr_debug("No more buffers in queue\n");
> +                       END_USE(vq);
> +                       return NULL;
> +               }
> +
> +               /*
> +                * Only get used array entries after they have been
> +                * exposed by host.
> +                */
> +               virtio_rmb(vq->weak_barriers);
> +
> +               vq->batch_last.id = virtio32_to_cpu(vq->vq.vdev,
> +                                   vq->split.vring.used->ring[last_used_idx].id);
> +               vq->batch_last.len = virtio32_to_cpu(vq->vq.vdev,
> +                                    vq->split.vring.used->ring[last_used_idx].len);
> +       }
> +
> +       if (vq->batch_last.id == last_used) {
> +               vq->batch_last.id = UINT_MAX;
> +               *len = vq->batch_last.len;
> +       } else {
> +               *len = vq->split.desc_state[last_used].total_len;
> +       }
> +
> +       if (unlikely(last_used >= num)) {
> +               BAD_RING(vq, "id %u out of range\n", last_used);
> +               return NULL;
> +       }
> +       if (unlikely(!vq->split.desc_state[last_used].data)) {
> +               BAD_RING(vq, "id %u is not a head!\n", last_used);
> +               return NULL;
> +       }
> +
> +       /* detach_buf_split clears data, so grab it now. */
> +       ret = vq->split.desc_state[last_used].data;
> +       detach_buf_split_in_order(vq, last_used, ctx);
> +
> +       vq->last_used_idx++;
> +       vq->last_used += (vq->vq.num_free - num_free);
> +       /* If we expect an interrupt for the next entry, tell host
> +        * by writing event index and flush out the write before
> +        * the read in the next get_buf call. */
> +       if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
> +               virtio_store_mb(vq->weak_barriers,
> +                               &vring_used_event(&vq->split.vring),
> +                               cpu_to_virtio16(vq->vq.vdev, vq->last_used_idx));
> +
> +       LAST_ADD_TIME_INVALID(vq);
> +
> +       END_USE(vq);
> +       return ret;
> +}
> +
>  static void virtqueue_disable_cb_split(struct vring_virtqueue *vq)
>  {
>         if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
> @@ -1012,7 +1135,10 @@ static void *virtqueue_detach_unused_buf_split(struct vring_virtqueue *vq)
>                         continue;
>                 /* detach_buf_split clears data, so grab it now. */
>                 buf = vq->split.desc_state[i].data;
> -               detach_buf_split(vq, i, NULL);
> +               if (virtqueue_is_in_order(vq))
> +                       detach_buf_split_in_order(vq, i, NULL);
> +               else
> +                       detach_buf_split(vq, i, NULL);
>                 vq->split.avail_idx_shadow--;
>                 vq->split.vring.avail->idx = cpu_to_virtio16(vq->vq.vdev,
>                                 vq->split.avail_idx_shadow);
> @@ -1075,6 +1201,7 @@ static void virtqueue_vring_attach_split(struct vring_virtqueue *vq,
>
>         /* Put everything in free lists. */
>         vq->free_head = 0;
> +       vq->batch_last.id = UINT_MAX;
>  }
>
>  static int vring_alloc_state_extra_split(struct vring_virtqueue_split *vring_split)
> @@ -1186,7 +1313,6 @@ static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
>         if (!vq)
>                 return NULL;
>
> -       vq->layout = SPLIT;
>         vq->vq.callback = callback;
>         vq->vq.vdev = vdev;
>         vq->vq.name = name;
> @@ -1206,6 +1332,8 @@ static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
>         vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
>                 !context;
>         vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
> +       vq->layout = virtio_has_feature(vdev, VIRTIO_F_IN_ORDER) ?
> +                    SPLIT_IN_ORDER : SPLIT;
>
>         if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
>                 vq->weak_barriers = false;
> @@ -1363,13 +1491,14 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
>                                          unsigned int in_sgs,
>                                          void *data,
>                                          bool premapped,
> -                                        gfp_t gfp)
> +                                        gfp_t gfp,
> +                                        u16 id)
>  {
>         struct vring_desc_extra *extra;
>         struct vring_packed_desc *desc;
>         struct scatterlist *sg;
> -       unsigned int i, n, err_idx, len;
> -       u16 head, id;
> +       unsigned int i, n, err_idx, len, total_len = 0;
> +       u16 head;
>         dma_addr_t addr;
>
>         head = vq->packed.next_avail_idx;
> @@ -1387,8 +1516,6 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
>         }
>
>         i = 0;
> -       id = vq->free_head;
> -       BUG_ON(id == vq->packed.vring.num);
>
>         for (n = 0; n < out_sgs + in_sgs; n++) {
>                 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
> @@ -1408,6 +1535,7 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
>                                 extra[i].flags = n < out_sgs ?  0 : VRING_DESC_F_WRITE;
>                         }
>
> +                       total_len += len;
>                         i++;
>                 }
>         }
> @@ -1454,13 +1582,15 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
>                                 1 << VRING_PACKED_DESC_F_USED;
>         }
>         vq->packed.next_avail_idx = n;
> -       vq->free_head = vq->packed.desc_extra[id].next;
> +       if (!virtqueue_is_in_order(vq))
> +               vq->free_head = vq->packed.desc_extra[id].next;
>
>         /* Store token and indirect buffer state. */
>         vq->packed.desc_state[id].num = 1;
>         vq->packed.desc_state[id].data = data;
>         vq->packed.desc_state[id].indir_desc = desc;
>         vq->packed.desc_state[id].last = id;
> +       vq->packed.desc_state[id].total_len = total_len;
>
>         vq->num_added += 1;
>
> @@ -1513,8 +1643,11 @@ static inline int virtqueue_add_packed(struct vring_virtqueue *vq,
>         BUG_ON(total_sg == 0);
>
>         if (virtqueue_use_indirect(vq, total_sg)) {
> +               id = vq->free_head;
> +               BUG_ON(id == vq->packed.vring.num);
>                 err = virtqueue_add_indirect_packed(vq, sgs, total_sg, out_sgs,
> -                                                   in_sgs, data, premapped, gfp);
> +                                                   in_sgs, data, premapped,
> +                                                   gfp, id);
>                 if (err != -ENOMEM) {
>                         END_USE(vq);
>                         return err;
> @@ -1635,6 +1768,156 @@ static inline int virtqueue_add_packed(struct vring_virtqueue *vq,
>         return -EIO;
>  }
>
> +static inline int virtqueue_add_packed_in_order(struct vring_virtqueue *vq,
> +                                               struct scatterlist *sgs[],
> +                                               unsigned int total_sg,
> +                                               unsigned int out_sgs,
> +                                               unsigned int in_sgs,
> +                                               void *data,
> +                                               void *ctx,
> +                                               bool premapped,
> +                                               gfp_t gfp)
> +{
> +       struct vring_packed_desc *desc;
> +       struct scatterlist *sg;
> +       unsigned int i, n, sg_count, err_idx, total_len = 0;
> +       __le16 head_flags, flags;
> +       u16 head, avail_used_flags;
> +       int err;
> +
> +       START_USE(vq);
> +
> +       BUG_ON(data == NULL);
> +       BUG_ON(ctx && vq->indirect);
> +
> +       if (unlikely(vq->broken)) {
> +               END_USE(vq);
> +               return -EIO;
> +       }
> +
> +       LAST_ADD_TIME_UPDATE(vq);
> +
> +       BUG_ON(total_sg == 0);
> +
> +       if (virtqueue_use_indirect(vq, total_sg)) {
> +               err = virtqueue_add_indirect_packed(vq, sgs, total_sg, out_sgs,
> +                                                   in_sgs, data, premapped, gfp,
> +                                                   vq->packed.next_avail_idx);
> +               if (err != -ENOMEM) {
> +                       END_USE(vq);
> +                       return err;
> +               }
> +
> +               /* fall back on direct */
> +       }
> +
> +       head = vq->packed.next_avail_idx;
> +       avail_used_flags = vq->packed.avail_used_flags;
> +
> +       WARN_ON_ONCE(total_sg > vq->packed.vring.num && !vq->indirect);
> +
> +       desc = vq->packed.vring.desc;
> +       i = head;
> +
> +       if (unlikely(vq->vq.num_free < total_sg)) {
> +               pr_debug("Can't add buf len %i - avail = %i\n",
> +                        total_sg, vq->vq.num_free);
> +               END_USE(vq);
> +               return -ENOSPC;
> +       }
> +
> +       sg_count = 0;
> +       for (n = 0; n < out_sgs + in_sgs; n++) {
> +               for (sg = sgs[n]; sg; sg = sg_next(sg)) {
> +                       dma_addr_t addr;
> +                       u32 len;
> +
> +                       flags = 0;
> +                       if (++sg_count != total_sg)
> +                               flags |= cpu_to_le16(VRING_DESC_F_NEXT);
> +                       if (n >= out_sgs)
> +                               flags |= cpu_to_le16(VRING_DESC_F_WRITE);
> +
> +                       if (vring_map_one_sg(vq, sg, n < out_sgs ?
> +                                            DMA_TO_DEVICE : DMA_FROM_DEVICE,
> +                                            &addr, &len, premapped))
> +                               goto unmap_release;
> +
> +                       flags |= cpu_to_le16(vq->packed.avail_used_flags);
> +
> +                       if (i == head)
> +                               head_flags = flags;
> +                       else
> +                               desc[i].flags = flags;
> +
> +                       desc[i].addr = cpu_to_le64(addr);
> +                       desc[i].len = cpu_to_le32(len);
> +                       desc[i].id = cpu_to_le16(head);
> +
> +                       if (unlikely(vq->use_map_api)) {
> +                               vq->packed.desc_extra[i].addr = premapped ?
> +                                     DMA_MAPPING_ERROR: addr;
> +                               vq->packed.desc_extra[i].len = len;
> +                               vq->packed.desc_extra[i].flags =
> +                                       le16_to_cpu(flags);
> +                       }
> +
> +                       if ((unlikely(++i >= vq->packed.vring.num))) {
> +                               i = 0;
> +                               vq->packed.avail_used_flags ^=
> +                                       1 << VRING_PACKED_DESC_F_AVAIL |
> +                                       1 << VRING_PACKED_DESC_F_USED;
> +                               vq->packed.avail_wrap_counter ^= 1;
> +                       }
> +
> +                       total_len += len;
> +               }
> +       }
> +
> +       /* We're using some buffers from the free list. */
> +       vq->vq.num_free -= total_sg;
> +
> +       /* Update free pointer */
> +       vq->packed.next_avail_idx = i;
> +
> +       /* Store token. */
> +       vq->packed.desc_state[head].num = total_sg;
> +       vq->packed.desc_state[head].data = data;
> +       vq->packed.desc_state[head].indir_desc = ctx;
> +       vq->packed.desc_state[head].total_len = total_len;
> +
> +       /*
> +        * A driver MUST NOT make the first descriptor in the list
> +        * available before all subsequent descriptors comprising
> +        * the list are made available.
> +        */
> +       virtio_wmb(vq->weak_barriers);
> +       vq->packed.vring.desc[head].flags = head_flags;
> +       vq->num_added += total_sg;
> +
> +       pr_debug("Added buffer head %i to %p\n", head, vq);
> +       END_USE(vq);
> +
> +       return 0;
> +
> +unmap_release:
> +       err_idx = i;
> +       i = head;
> +       vq->packed.avail_used_flags = avail_used_flags;
> +
> +       for (n = 0; n < total_sg; n++) {
> +               if (i == err_idx)
> +                       break;
> +               vring_unmap_extra_packed(vq, &vq->packed.desc_extra[i]);
> +               i++;
> +               if (i >= vq->packed.vring.num)
> +                       i = 0;
> +       }
> +
> +       END_USE(vq);
> +       return -EIO;
> +}

This function duplicates a lot of logic from virtqueue_add_packed in
my opinion. To make descriptors available is the same process in
order, except for saving total_len. I like this new version more, as
it gets rid of descs_used, that is always = total_sg. Maybe it is
better to just perform these steps conditionally, the same way as
virtqueue_add_split?

> +
>  static bool virtqueue_kick_prepare_packed(struct vring_virtqueue *vq)
>  {
>         u16 new, old, off_wrap, flags, wrap_counter, event_idx;
> @@ -1796,10 +2079,82 @@ static void update_last_used_idx_packed(struct vring_virtqueue *vq,
>                                 cpu_to_le16(vq->last_used_idx));
>  }
>
> +static bool more_used_packed_in_order(const struct vring_virtqueue *vq)
> +{
> +       if (vq->batch_last.id != UINT_MAX)
> +               return true;
> +
> +       return virtqueue_poll_packed(vq, READ_ONCE(vq->last_used_idx));
> +}
> +
> +static void *virtqueue_get_buf_ctx_packed_in_order(struct vring_virtqueue *vq,
> +                                                  unsigned int *len,
> +                                                  void **ctx)
> +{
> +       unsigned int num = vq->packed.vring.num;
> +       u16 last_used, last_used_idx;
> +       bool used_wrap_counter;
> +       void *ret;
> +
> +       START_USE(vq);
> +
> +       if (unlikely(vq->broken)) {
> +               END_USE(vq);
> +               return NULL;
> +       }
> +
> +       last_used_idx = vq->last_used_idx;
> +       used_wrap_counter = packed_used_wrap_counter(last_used_idx);
> +       last_used = packed_last_used(last_used_idx);
> +
> +       if (vq->batch_last.id == UINT_MAX) {
> +               if (!more_used_packed_in_order(vq)) {
> +                       pr_debug("No more buffers in queue\n");
> +                       END_USE(vq);
> +                       return NULL;
> +               }
> +               /* Only get used elements after they have been exposed by host. */
> +               virtio_rmb(vq->weak_barriers);
> +               vq->batch_last.id =
> +                       le16_to_cpu(vq->packed.vring.desc[last_used].id);
> +               vq->batch_last.len =
> +                       le32_to_cpu(vq->packed.vring.desc[last_used].len);
> +       }
> +
> +       if (vq->batch_last.id == last_used) {
> +               vq->batch_last.id = UINT_MAX;
> +               *len = vq->batch_last.len;
> +       } else {
> +               *len = vq->packed.desc_state[last_used].total_len;
> +       }
> +
> +       if (unlikely(last_used >= num)) {
> +               BAD_RING(vq, "id %u out of range\n", last_used);
> +               return NULL;
> +       }
> +       if (unlikely(!vq->packed.desc_state[last_used].data)) {
> +               BAD_RING(vq, "id %u is not a head!\n", last_used);
> +               return NULL;
> +       }
> +
> +       /* detach_buf_packed clears data, so grab it now. */
> +       ret = vq->packed.desc_state[last_used].data;
> +       detach_buf_packed_in_order(vq, last_used, ctx);
> +
> +       update_last_used_idx_packed(vq, last_used, last_used,
> +                                   used_wrap_counter);
> +
> +       LAST_ADD_TIME_INVALID(vq);
> +
> +       END_USE(vq);
> +       return ret;
> +}
> +
>  static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
>                                           unsigned int *len,
>                                           void **ctx)
>  {
> +       unsigned int num = vq->packed.vring.num;
>         u16 last_used, id, last_used_idx;
>         bool used_wrap_counter;
>         void *ret;
> @@ -1826,7 +2181,7 @@ static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
>         id = le16_to_cpu(vq->packed.vring.desc[last_used].id);
>         *len = le32_to_cpu(vq->packed.vring.desc[last_used].len);
>
> -       if (unlikely(id >= vq->packed.vring.num)) {
> +       if (unlikely(id >= num)) {
>                 BAD_RING(vq, "id %u out of range\n", id);
>                 return NULL;
>         }
> @@ -1967,7 +2322,10 @@ static void *virtqueue_detach_unused_buf_packed(struct vring_virtqueue *vq)
>                         continue;
>                 /* detach_buf clears data, so grab it now. */
>                 buf = vq->packed.desc_state[i].data;
> -               detach_buf_packed(vq, i, NULL);
> +               if (virtqueue_is_in_order(vq))
> +                       detach_buf_packed_in_order(vq, i, NULL);
> +               else
> +                       detach_buf_packed(vq, i, NULL);
>                 END_USE(vq);
>                 return buf;
>         }
> @@ -1993,6 +2351,8 @@ static struct vring_desc_extra *vring_alloc_desc_extra(unsigned int num)
>         for (i = 0; i < num - 1; i++)
>                 desc_extra[i].next = i + 1;
>
> +       desc_extra[num - 1].next = 0;
> +
>         return desc_extra;
>  }
>
> @@ -2124,10 +2484,17 @@ static void virtqueue_vring_attach_packed(struct vring_virtqueue *vq,
>  {
>         vq->packed = *vring_packed;
>
> -       /* Put everything in free lists. */
> -       vq->free_head = 0;
> +       if (virtqueue_is_in_order(vq)) {
> +               vq->batch_last.id = UINT_MAX;
> +       } else {
> +               /*
> +                * Put everything in free lists. Note that
> +                * next_avail_idx is sufficient with IN_ORDER so
> +                * free_head is unused.
> +                */
> +               vq->free_head = 0;
> +       }
>  }
> -
>  static void virtqueue_reset_packed(struct vring_virtqueue *vq)
>  {
>         memset(vq->packed.vring.device, 0, vq->packed.event_size_in_bytes);
> @@ -2172,13 +2539,14 @@ static struct virtqueue *__vring_new_virtqueue_packed(unsigned int index,
>  #else
>         vq->broken = false;
>  #endif
> -       vq->layout = PACKED;
>         vq->map = map;
>         vq->use_map_api = vring_use_map_api(vdev);
>
>         vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
>                 !context;
>         vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
> +       vq->layout = virtio_has_feature(vdev, VIRTIO_F_IN_ORDER) ?
> +                    PACKED_IN_ORDER : PACKED;
>
>         if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
>                 vq->weak_barriers = false;
> @@ -2288,9 +2656,39 @@ static const struct virtqueue_ops packed_ops = {
>         .reset = virtqueue_reset_packed,
>  };
>
> +static const struct virtqueue_ops split_in_order_ops = {
> +       .add = virtqueue_add_split,
> +       .get = virtqueue_get_buf_ctx_split_in_order,
> +       .kick_prepare = virtqueue_kick_prepare_split,
> +       .disable_cb = virtqueue_disable_cb_split,
> +       .enable_cb_delayed = virtqueue_enable_cb_delayed_split,
> +       .enable_cb_prepare = virtqueue_enable_cb_prepare_split,
> +       .poll = virtqueue_poll_split,
> +       .detach_unused_buf = virtqueue_detach_unused_buf_split,
> +       .more_used = more_used_split_in_order,
> +       .resize = virtqueue_resize_split,
> +       .reset = virtqueue_reset_split,
> +};
> +
> +static const struct virtqueue_ops packed_in_order_ops = {
> +       .add = virtqueue_add_packed_in_order,
> +       .get = virtqueue_get_buf_ctx_packed_in_order,
> +       .kick_prepare = virtqueue_kick_prepare_packed,
> +       .disable_cb = virtqueue_disable_cb_packed,
> +       .enable_cb_delayed = virtqueue_enable_cb_delayed_packed,
> +       .enable_cb_prepare = virtqueue_enable_cb_prepare_packed,
> +       .poll = virtqueue_poll_packed,
> +       .detach_unused_buf = virtqueue_detach_unused_buf_packed,
> +       .more_used = more_used_packed_in_order,
> +       .resize = virtqueue_resize_packed,
> +       .reset = virtqueue_reset_packed,
> +};
> +
>  static const struct virtqueue_ops *const all_ops[VQ_TYPE_MAX] = {
>         [SPLIT] = &split_ops,
> -       [PACKED] = &packed_ops
> +       [PACKED] = &packed_ops,
> +       [SPLIT_IN_ORDER] = &split_in_order_ops,
> +       [PACKED_IN_ORDER] = &packed_in_order_ops,
>  };
>
>  static int virtqueue_disable_and_recycle(struct virtqueue *_vq,
> @@ -2346,6 +2744,12 @@ static int virtqueue_enable_after_reset(struct virtqueue *_vq)
>         case PACKED:                                                    \
>                 ret = all_ops[PACKED]->op(vq, ##__VA_ARGS__);           \
>                 break;                                                  \
> +       case SPLIT_IN_ORDER:                                            \
> +               ret = all_ops[SPLIT_IN_ORDER]->op(vq, ##__VA_ARGS__);   \
> +               break;                                                  \
> +       case PACKED_IN_ORDER:                                           \
> +               ret = all_ops[PACKED_IN_ORDER]->op(vq, ##__VA_ARGS__);  \
> +               break;                                                  \
>         default:                                                        \
>                 BUG();                                                  \
>                 break;                                                  \
> @@ -2362,10 +2766,16 @@ static int virtqueue_enable_after_reset(struct virtqueue *_vq)
>         case PACKED:                                    \
>                 all_ops[PACKED]->op(vq, ##__VA_ARGS__); \
>                 break;                                  \
> -       default:                                        \
> -               BUG();                                  \
> -               break;                                  \
> -       }                                               \
> +       case SPLIT_IN_ORDER:                                            \
> +               all_ops[SPLIT_IN_ORDER]->op(vq, ##__VA_ARGS__); \
> +               break;                                                  \
> +       case PACKED_IN_ORDER:                                           \
> +               all_ops[PACKED_IN_ORDER]->op(vq, ##__VA_ARGS__);        \
> +               break;                                                  \
> +       default:                                                        \
> +               BUG();                                                  \
> +               break;                                                  \
> +       }                                                               \
>  })
>
>  static inline int virtqueue_add(struct virtqueue *_vq,
> @@ -3081,6 +3491,8 @@ void vring_transport_features(struct virtio_device *vdev)
>                         break;
>                 case VIRTIO_F_NOTIFICATION_DATA:
>                         break;
> +               case VIRTIO_F_IN_ORDER:
> +                       break;
>                 default:
>                         /* We don't understand this bit. */
>                         __virtio_clear_bit(vdev, i);
> --
> 2.31.1
>

Re: [PATCH V9 19/19] virtio_ring: add in order support

[Eugenio Perez Martin]

On Fri, Dec 19, 2025 at 11:44 AM Eugenio Perez Martin
<eperezma@redhat.com> wrote:
>
> On Tue, Nov 25, 2025 at 8:10 AM Jason Wang <jasowang@redhat.com> wrote:
> >
> > This patch implements in order support for both split virtqueue and
> > packed virtqueue. Performance could be gained for the device where the
> > memory access could be expensive (e.g vhost-net or a real PCI device):
> >
> > Benchmark with KVM guest:
> >
> > Vhost-net on the host: (pktgen + XDP_DROP):
> >
> >          in_order=off | in_order=on | +%
> >     TX:  4.51Mpps     | 5.30Mpps    | +17%
> >     RX:  3.47Mpps     | 3.61Mpps    | + 4%
> >
> > Vhost-user(testpmd) on the host: (pktgen/XDP_DROP):
> >
> > For split virtqueue:
> >
> >          in_order=off | in_order=on | +%
> >     TX:  5.60Mpps     | 5.60Mpps    | +0.0%
> >     RX:  9.16Mpps     | 9.61Mpps    | +4.9%
> >
> > For packed virtqueue:
> >
> >          in_order=off | in_order=on | +%
> >     TX:  5.60Mpps     | 5.70Mpps    | +1.7%
> >     RX:  10.6Mpps     | 10.8Mpps    | +1.8%
> >
> > Benchmark also shows no performance impact for in_order=off for queue
> > size with 256 and 1024.
> >
> > Signed-off-by: Jason Wang <jasowang@redhat.com>
>
> Reviewed-by: Eugenio Pérez <eperezma@redhat.com>
>
> I have one nitpick though, but it can be added on top or just ignored.
>
> > ---
> >  drivers/virtio/virtio_ring.c | 458 +++++++++++++++++++++++++++++++++--
> >  1 file changed, 435 insertions(+), 23 deletions(-)
> >
> > diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> > index 61884e031b94..9847b23f82ea 100644
> > --- a/drivers/virtio/virtio_ring.c
> > +++ b/drivers/virtio/virtio_ring.c
> > @@ -70,6 +70,8 @@
> >  enum vq_layout {
> >         SPLIT = 0,
> >         PACKED,
> > +       SPLIT_IN_ORDER,
> > +       PACKED_IN_ORDER,
> >         VQ_TYPE_MAX,
> >  };
> >
> > @@ -80,6 +82,7 @@ struct vring_desc_state_split {
> >          * allocated together. So we won't stress more to the memory allocator.
> >          */
> >         struct vring_desc *indir_desc;
> > +       u32 total_len;                  /* Buffer Length */
> >  };
> >
> >  struct vring_desc_state_packed {
> > @@ -91,6 +94,7 @@ struct vring_desc_state_packed {
> >         struct vring_packed_desc *indir_desc;
> >         u16 num;                        /* Descriptor list length. */
> >         u16 last;                       /* The last desc state in a list. */
> > +       u32 total_len;                  /* Buffer Length */
> >  };
> >
> >  struct vring_desc_extra {
> > @@ -205,8 +209,24 @@ struct vring_virtqueue {
> >
> >         enum vq_layout layout;
> >
> > -       /* Head of free buffer list. */
> > +       /*
> > +        * Without IN_ORDER it's the head of free buffer list. With
> > +        * IN_ORDER and SPLIT, it's the next available buffer
> > +        * index. With IN_ORDER and PACKED, it's unused.
> > +        */
> >         unsigned int free_head;
> > +
> > +       /*
> > +        * With IN_ORDER, once we see an in-order batch, this stores
> > +        * this last entry, and until we return the last buffer.
> > +        * After this, id is set to vq.num to mark it invalid.
> > +        * Unused without IN_ORDER.
> > +        */
> > +       struct used_entry {
> > +               u32 id;
> > +               u32 len;
> > +       } batch_last;
> > +
> >         /* Number we've added since last sync. */
> >         unsigned int num_added;
> >
> > @@ -218,6 +238,11 @@ struct vring_virtqueue {
> >          */
> >         u16 last_used_idx;
> >
> > +       /* With IN_ORDER and SPLIT, last descriptor id we used to
> > +        * detach buffer.
> > +        */
> > +       u16 last_used;
> > +
> >         /* Hint for event idx: already triggered no need to disable. */
> >         bool event_triggered;
> >
> > @@ -259,7 +284,12 @@ static void vring_free(struct virtqueue *_vq);
> >
> >  static inline bool virtqueue_is_packed(const struct vring_virtqueue *vq)
> >  {
> > -       return vq->layout == PACKED;
> > +       return vq->layout == PACKED || vq->layout == PACKED_IN_ORDER;
> > +}
> > +
> > +static inline bool virtqueue_is_in_order(const struct vring_virtqueue *vq)
> > +{
> > +       return vq->layout == SPLIT_IN_ORDER || vq->layout == PACKED_IN_ORDER;
> >  }
> >
> >  static bool virtqueue_use_indirect(const struct vring_virtqueue *vq,
> > @@ -469,6 +499,8 @@ static void virtqueue_init(struct vring_virtqueue *vq, u32 num)
> >         else
> >                 vq->last_used_idx = 0;
> >
> > +       vq->last_used = 0;
> > +
> >         vq->event_triggered = false;
> >         vq->num_added = 0;
> >
> > @@ -576,6 +608,8 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
> >         struct scatterlist *sg;
> >         struct vring_desc *desc;
> >         unsigned int i, n, avail, descs_used, err_idx, sg_count = 0;
> > +       /* Total length for in-order */
> > +       unsigned int total_len = 0;
> >         int head;
> >         bool indirect;
> >
> > @@ -648,6 +682,7 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
> >                          */
> >                         i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
> >                                                      len, flags ,premapped);
> > +                       total_len += len;
> >                 }
> >         }
> >         for (; n < (out_sgs + in_sgs); n++) {
> > @@ -667,6 +702,7 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
> >                          */
> >                         i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
> >                                                      len, flags, premapped);
> > +                       total_len += len;
> >                 }
> >         }
> >
> > @@ -689,7 +725,12 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
> >         vq->vq.num_free -= descs_used;
> >
> >         /* Update free pointer */
> > -       if (indirect)
> > +       if (virtqueue_is_in_order(vq)) {
> > +               vq->free_head += descs_used;
> > +               if (vq->free_head >= vq->split.vring.num)
> > +                       vq->free_head -= vq->split.vring.num;
> > +               vq->split.desc_state[head].total_len = total_len;
> > +       } else if (indirect)
> >                 vq->free_head = vq->split.desc_extra[head].next;
> >         else
> >                 vq->free_head = i;
> > @@ -862,6 +903,14 @@ static bool more_used_split(const struct vring_virtqueue *vq)
> >         return virtqueue_poll_split(vq, vq->last_used_idx);
> >  }
> >
> > +static bool more_used_split_in_order(const struct vring_virtqueue *vq)
> > +{
> > +       if (vq->batch_last.id != UINT_MAX)
> > +               return true;
> > +
> > +       return virtqueue_poll_split(vq, vq->last_used_idx);
> > +}
> > +
> >  static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
> >                                          unsigned int *len,
> >                                          void **ctx)
> > @@ -919,6 +968,80 @@ static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
> >         return ret;
> >  }
> >
> > +static void *virtqueue_get_buf_ctx_split_in_order(struct vring_virtqueue *vq,
> > +                                                 unsigned int *len,
> > +                                                 void **ctx)
> > +{
> > +       void *ret;
> > +       unsigned int num = vq->split.vring.num;
> > +       unsigned int num_free = vq->vq.num_free;
> > +       u16 last_used, last_used_idx;
> > +
> > +       START_USE(vq);
> > +
> > +       if (unlikely(vq->broken)) {
> > +               END_USE(vq);
> > +               return NULL;
> > +       }
> > +
> > +       last_used = vq->last_used & (num - 1);
> > +       last_used_idx = vq->last_used_idx & (num - 1);
> > +
> > +       if (vq->batch_last.id == UINT_MAX) {
> > +               if (!more_used_split_in_order(vq)) {
> > +                       pr_debug("No more buffers in queue\n");
> > +                       END_USE(vq);
> > +                       return NULL;
> > +               }
> > +
> > +               /*
> > +                * Only get used array entries after they have been
> > +                * exposed by host.
> > +                */
> > +               virtio_rmb(vq->weak_barriers);
> > +
> > +               vq->batch_last.id = virtio32_to_cpu(vq->vq.vdev,
> > +                                   vq->split.vring.used->ring[last_used_idx].id);
> > +               vq->batch_last.len = virtio32_to_cpu(vq->vq.vdev,
> > +                                    vq->split.vring.used->ring[last_used_idx].len);
> > +       }
> > +
> > +       if (vq->batch_last.id == last_used) {
> > +               vq->batch_last.id = UINT_MAX;
> > +               *len = vq->batch_last.len;
> > +       } else {
> > +               *len = vq->split.desc_state[last_used].total_len;
> > +       }
> > +
> > +       if (unlikely(last_used >= num)) {
> > +               BAD_RING(vq, "id %u out of range\n", last_used);
> > +               return NULL;
> > +       }
> > +       if (unlikely(!vq->split.desc_state[last_used].data)) {
> > +               BAD_RING(vq, "id %u is not a head!\n", last_used);
> > +               return NULL;
> > +       }
> > +
> > +       /* detach_buf_split clears data, so grab it now. */
> > +       ret = vq->split.desc_state[last_used].data;
> > +       detach_buf_split_in_order(vq, last_used, ctx);
> > +
> > +       vq->last_used_idx++;
> > +       vq->last_used += (vq->vq.num_free - num_free);
> > +       /* If we expect an interrupt for the next entry, tell host
> > +        * by writing event index and flush out the write before
> > +        * the read in the next get_buf call. */
> > +       if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
> > +               virtio_store_mb(vq->weak_barriers,
> > +                               &vring_used_event(&vq->split.vring),
> > +                               cpu_to_virtio16(vq->vq.vdev, vq->last_used_idx));
> > +
> > +       LAST_ADD_TIME_INVALID(vq);
> > +
> > +       END_USE(vq);
> > +       return ret;
> > +}
> > +
> >  static void virtqueue_disable_cb_split(struct vring_virtqueue *vq)
> >  {
> >         if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
> > @@ -1012,7 +1135,10 @@ static void *virtqueue_detach_unused_buf_split(struct vring_virtqueue *vq)
> >                         continue;
> >                 /* detach_buf_split clears data, so grab it now. */
> >                 buf = vq->split.desc_state[i].data;
> > -               detach_buf_split(vq, i, NULL);
> > +               if (virtqueue_is_in_order(vq))
> > +                       detach_buf_split_in_order(vq, i, NULL);
> > +               else
> > +                       detach_buf_split(vq, i, NULL);
> >                 vq->split.avail_idx_shadow--;
> >                 vq->split.vring.avail->idx = cpu_to_virtio16(vq->vq.vdev,
> >                                 vq->split.avail_idx_shadow);
> > @@ -1075,6 +1201,7 @@ static void virtqueue_vring_attach_split(struct vring_virtqueue *vq,
> >
> >         /* Put everything in free lists. */
> >         vq->free_head = 0;
> > +       vq->batch_last.id = UINT_MAX;
> >  }
> >
> >  static int vring_alloc_state_extra_split(struct vring_virtqueue_split *vring_split)
> > @@ -1186,7 +1313,6 @@ static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
> >         if (!vq)
> >                 return NULL;
> >
> > -       vq->layout = SPLIT;
> >         vq->vq.callback = callback;
> >         vq->vq.vdev = vdev;
> >         vq->vq.name = name;
> > @@ -1206,6 +1332,8 @@ static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
> >         vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
> >                 !context;
> >         vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
> > +       vq->layout = virtio_has_feature(vdev, VIRTIO_F_IN_ORDER) ?
> > +                    SPLIT_IN_ORDER : SPLIT;
> >
> >         if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
> >                 vq->weak_barriers = false;
> > @@ -1363,13 +1491,14 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
> >                                          unsigned int in_sgs,
> >                                          void *data,
> >                                          bool premapped,
> > -                                        gfp_t gfp)
> > +                                        gfp_t gfp,
> > +                                        u16 id)
> >  {
> >         struct vring_desc_extra *extra;
> >         struct vring_packed_desc *desc;
> >         struct scatterlist *sg;
> > -       unsigned int i, n, err_idx, len;
> > -       u16 head, id;
> > +       unsigned int i, n, err_idx, len, total_len = 0;
> > +       u16 head;
> >         dma_addr_t addr;
> >
> >         head = vq->packed.next_avail_idx;
> > @@ -1387,8 +1516,6 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
> >         }
> >
> >         i = 0;
> > -       id = vq->free_head;
> > -       BUG_ON(id == vq->packed.vring.num);
> >
> >         for (n = 0; n < out_sgs + in_sgs; n++) {
> >                 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
> > @@ -1408,6 +1535,7 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
> >                                 extra[i].flags = n < out_sgs ?  0 : VRING_DESC_F_WRITE;
> >                         }
> >
> > +                       total_len += len;
> >                         i++;
> >                 }
> >         }
> > @@ -1454,13 +1582,15 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
> >                                 1 << VRING_PACKED_DESC_F_USED;
> >         }
> >         vq->packed.next_avail_idx = n;
> > -       vq->free_head = vq->packed.desc_extra[id].next;
> > +       if (!virtqueue_is_in_order(vq))
> > +               vq->free_head = vq->packed.desc_extra[id].next;
> >
> >         /* Store token and indirect buffer state. */
> >         vq->packed.desc_state[id].num = 1;
> >         vq->packed.desc_state[id].data = data;
> >         vq->packed.desc_state[id].indir_desc = desc;
> >         vq->packed.desc_state[id].last = id;
> > +       vq->packed.desc_state[id].total_len = total_len;
> >
> >         vq->num_added += 1;
> >
> > @@ -1513,8 +1643,11 @@ static inline int virtqueue_add_packed(struct vring_virtqueue *vq,
> >         BUG_ON(total_sg == 0);
> >
> >         if (virtqueue_use_indirect(vq, total_sg)) {
> > +               id = vq->free_head;
> > +               BUG_ON(id == vq->packed.vring.num);
> >                 err = virtqueue_add_indirect_packed(vq, sgs, total_sg, out_sgs,
> > -                                                   in_sgs, data, premapped, gfp);
> > +                                                   in_sgs, data, premapped,
> > +                                                   gfp, id);
> >                 if (err != -ENOMEM) {
> >                         END_USE(vq);
> >                         return err;
> > @@ -1635,6 +1768,156 @@ static inline int virtqueue_add_packed(struct vring_virtqueue *vq,
> >         return -EIO;
> >  }
> >
> > +static inline int virtqueue_add_packed_in_order(struct vring_virtqueue *vq,
> > +                                               struct scatterlist *sgs[],
> > +                                               unsigned int total_sg,
> > +                                               unsigned int out_sgs,
> > +                                               unsigned int in_sgs,
> > +                                               void *data,
> > +                                               void *ctx,
> > +                                               bool premapped,
> > +                                               gfp_t gfp)
> > +{
> > +       struct vring_packed_desc *desc;
> > +       struct scatterlist *sg;
> > +       unsigned int i, n, sg_count, err_idx, total_len = 0;
> > +       __le16 head_flags, flags;
> > +       u16 head, avail_used_flags;
> > +       int err;
> > +
> > +       START_USE(vq);
> > +
> > +       BUG_ON(data == NULL);
> > +       BUG_ON(ctx && vq->indirect);
> > +
> > +       if (unlikely(vq->broken)) {
> > +               END_USE(vq);
> > +               return -EIO;
> > +       }
> > +
> > +       LAST_ADD_TIME_UPDATE(vq);
> > +
> > +       BUG_ON(total_sg == 0);
> > +
> > +       if (virtqueue_use_indirect(vq, total_sg)) {
> > +               err = virtqueue_add_indirect_packed(vq, sgs, total_sg, out_sgs,
> > +                                                   in_sgs, data, premapped, gfp,
> > +                                                   vq->packed.next_avail_idx);
> > +               if (err != -ENOMEM) {
> > +                       END_USE(vq);
> > +                       return err;
> > +               }
> > +
> > +               /* fall back on direct */
> > +       }
> > +
> > +       head = vq->packed.next_avail_idx;
> > +       avail_used_flags = vq->packed.avail_used_flags;
> > +
> > +       WARN_ON_ONCE(total_sg > vq->packed.vring.num && !vq->indirect);
> > +
> > +       desc = vq->packed.vring.desc;
> > +       i = head;
> > +
> > +       if (unlikely(vq->vq.num_free < total_sg)) {
> > +               pr_debug("Can't add buf len %i - avail = %i\n",
> > +                        total_sg, vq->vq.num_free);
> > +               END_USE(vq);
> > +               return -ENOSPC;
> > +       }
> > +
> > +       sg_count = 0;
> > +       for (n = 0; n < out_sgs + in_sgs; n++) {
> > +               for (sg = sgs[n]; sg; sg = sg_next(sg)) {
> > +                       dma_addr_t addr;
> > +                       u32 len;
> > +
> > +                       flags = 0;
> > +                       if (++sg_count != total_sg)
> > +                               flags |= cpu_to_le16(VRING_DESC_F_NEXT);
> > +                       if (n >= out_sgs)
> > +                               flags |= cpu_to_le16(VRING_DESC_F_WRITE);
> > +
> > +                       if (vring_map_one_sg(vq, sg, n < out_sgs ?
> > +                                            DMA_TO_DEVICE : DMA_FROM_DEVICE,
> > +                                            &addr, &len, premapped))
> > +                               goto unmap_release;
> > +
> > +                       flags |= cpu_to_le16(vq->packed.avail_used_flags);
> > +
> > +                       if (i == head)
> > +                               head_flags = flags;
> > +                       else
> > +                               desc[i].flags = flags;
> > +
> > +                       desc[i].addr = cpu_to_le64(addr);
> > +                       desc[i].len = cpu_to_le32(len);
> > +                       desc[i].id = cpu_to_le16(head);
> > +
> > +                       if (unlikely(vq->use_map_api)) {
> > +                               vq->packed.desc_extra[i].addr = premapped ?
> > +                                     DMA_MAPPING_ERROR: addr;
> > +                               vq->packed.desc_extra[i].len = len;
> > +                               vq->packed.desc_extra[i].flags =
> > +                                       le16_to_cpu(flags);
> > +                       }
> > +
> > +                       if ((unlikely(++i >= vq->packed.vring.num))) {
> > +                               i = 0;
> > +                               vq->packed.avail_used_flags ^=
> > +                                       1 << VRING_PACKED_DESC_F_AVAIL |
> > +                                       1 << VRING_PACKED_DESC_F_USED;
> > +                               vq->packed.avail_wrap_counter ^= 1;
> > +                       }
> > +
> > +                       total_len += len;

I hit send too fast :). The used descriptor "length" field is the
amount of bytes written by the device. So total_len should only be
updated in the in descs, not in the out ones. Same with split. Maybe
rename to total_in_len is also useful, I'm calling that way in SVQ.

> > +               }
> > +       }
> > +
> > +       /* We're using some buffers from the free list. */
> > +       vq->vq.num_free -= total_sg;
> > +
> > +       /* Update free pointer */
> > +       vq->packed.next_avail_idx = i;
> > +
> > +       /* Store token. */
> > +       vq->packed.desc_state[head].num = total_sg;
> > +       vq->packed.desc_state[head].data = data;
> > +       vq->packed.desc_state[head].indir_desc = ctx;
> > +       vq->packed.desc_state[head].total_len = total_len;
> > +
> > +       /*
> > +        * A driver MUST NOT make the first descriptor in the list
> > +        * available before all subsequent descriptors comprising
> > +        * the list are made available.
> > +        */
> > +       virtio_wmb(vq->weak_barriers);
> > +       vq->packed.vring.desc[head].flags = head_flags;
> > +       vq->num_added += total_sg;
> > +
> > +       pr_debug("Added buffer head %i to %p\n", head, vq);
> > +       END_USE(vq);
> > +
> > +       return 0;
> > +
> > +unmap_release:
> > +       err_idx = i;
> > +       i = head;
> > +       vq->packed.avail_used_flags = avail_used_flags;
> > +
> > +       for (n = 0; n < total_sg; n++) {
> > +               if (i == err_idx)
> > +                       break;
> > +               vring_unmap_extra_packed(vq, &vq->packed.desc_extra[i]);
> > +               i++;
> > +               if (i >= vq->packed.vring.num)
> > +                       i = 0;
> > +       }
> > +
> > +       END_USE(vq);
> > +       return -EIO;
> > +}
>
> This function duplicates a lot of logic from virtqueue_add_packed in
> my opinion. To make descriptors available is the same process in
> order, except for saving total_len. I like this new version more, as
> it gets rid of descs_used, that is always = total_sg. Maybe it is
> better to just perform these steps conditionally, the same way as
> virtqueue_add_split?
>
> > +
> >  static bool virtqueue_kick_prepare_packed(struct vring_virtqueue *vq)
> >  {
> >         u16 new, old, off_wrap, flags, wrap_counter, event_idx;
> > @@ -1796,10 +2079,82 @@ static void update_last_used_idx_packed(struct vring_virtqueue *vq,
> >                                 cpu_to_le16(vq->last_used_idx));
> >  }
> >
> > +static bool more_used_packed_in_order(const struct vring_virtqueue *vq)
> > +{
> > +       if (vq->batch_last.id != UINT_MAX)
> > +               return true;
> > +
> > +       return virtqueue_poll_packed(vq, READ_ONCE(vq->last_used_idx));
> > +}
> > +
> > +static void *virtqueue_get_buf_ctx_packed_in_order(struct vring_virtqueue *vq,
> > +                                                  unsigned int *len,
> > +                                                  void **ctx)
> > +{
> > +       unsigned int num = vq->packed.vring.num;
> > +       u16 last_used, last_used_idx;
> > +       bool used_wrap_counter;
> > +       void *ret;
> > +
> > +       START_USE(vq);
> > +
> > +       if (unlikely(vq->broken)) {
> > +               END_USE(vq);
> > +               return NULL;
> > +       }
> > +
> > +       last_used_idx = vq->last_used_idx;
> > +       used_wrap_counter = packed_used_wrap_counter(last_used_idx);
> > +       last_used = packed_last_used(last_used_idx);
> > +
> > +       if (vq->batch_last.id == UINT_MAX) {
> > +               if (!more_used_packed_in_order(vq)) {
> > +                       pr_debug("No more buffers in queue\n");
> > +                       END_USE(vq);
> > +                       return NULL;
> > +               }
> > +               /* Only get used elements after they have been exposed by host. */
> > +               virtio_rmb(vq->weak_barriers);
> > +               vq->batch_last.id =
> > +                       le16_to_cpu(vq->packed.vring.desc[last_used].id);
> > +               vq->batch_last.len =
> > +                       le32_to_cpu(vq->packed.vring.desc[last_used].len);
> > +       }
> > +
> > +       if (vq->batch_last.id == last_used) {
> > +               vq->batch_last.id = UINT_MAX;
> > +               *len = vq->batch_last.len;
> > +       } else {
> > +               *len = vq->packed.desc_state[last_used].total_len;
> > +       }
> > +
> > +       if (unlikely(last_used >= num)) {
> > +               BAD_RING(vq, "id %u out of range\n", last_used);
> > +               return NULL;
> > +       }
> > +       if (unlikely(!vq->packed.desc_state[last_used].data)) {
> > +               BAD_RING(vq, "id %u is not a head!\n", last_used);
> > +               return NULL;
> > +       }
> > +
> > +       /* detach_buf_packed clears data, so grab it now. */
> > +       ret = vq->packed.desc_state[last_used].data;
> > +       detach_buf_packed_in_order(vq, last_used, ctx);
> > +
> > +       update_last_used_idx_packed(vq, last_used, last_used,
> > +                                   used_wrap_counter);
> > +
> > +       LAST_ADD_TIME_INVALID(vq);
> > +
> > +       END_USE(vq);
> > +       return ret;
> > +}
> > +
> >  static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
> >                                           unsigned int *len,
> >                                           void **ctx)
> >  {
> > +       unsigned int num = vq->packed.vring.num;
> >         u16 last_used, id, last_used_idx;
> >         bool used_wrap_counter;
> >         void *ret;
> > @@ -1826,7 +2181,7 @@ static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
> >         id = le16_to_cpu(vq->packed.vring.desc[last_used].id);
> >         *len = le32_to_cpu(vq->packed.vring.desc[last_used].len);
> >
> > -       if (unlikely(id >= vq->packed.vring.num)) {
> > +       if (unlikely(id >= num)) {
> >                 BAD_RING(vq, "id %u out of range\n", id);
> >                 return NULL;
> >         }
> > @@ -1967,7 +2322,10 @@ static void *virtqueue_detach_unused_buf_packed(struct vring_virtqueue *vq)
> >                         continue;
> >                 /* detach_buf clears data, so grab it now. */
> >                 buf = vq->packed.desc_state[i].data;
> > -               detach_buf_packed(vq, i, NULL);
> > +               if (virtqueue_is_in_order(vq))
> > +                       detach_buf_packed_in_order(vq, i, NULL);
> > +               else
> > +                       detach_buf_packed(vq, i, NULL);
> >                 END_USE(vq);
> >                 return buf;
> >         }
> > @@ -1993,6 +2351,8 @@ static struct vring_desc_extra *vring_alloc_desc_extra(unsigned int num)
> >         for (i = 0; i < num - 1; i++)
> >                 desc_extra[i].next = i + 1;
> >
> > +       desc_extra[num - 1].next = 0;
> > +
> >         return desc_extra;
> >  }
> >
> > @@ -2124,10 +2484,17 @@ static void virtqueue_vring_attach_packed(struct vring_virtqueue *vq,
> >  {
> >         vq->packed = *vring_packed;
> >
> > -       /* Put everything in free lists. */
> > -       vq->free_head = 0;
> > +       if (virtqueue_is_in_order(vq)) {
> > +               vq->batch_last.id = UINT_MAX;
> > +       } else {
> > +               /*
> > +                * Put everything in free lists. Note that
> > +                * next_avail_idx is sufficient with IN_ORDER so
> > +                * free_head is unused.
> > +                */
> > +               vq->free_head = 0;
> > +       }
> >  }
> > -
> >  static void virtqueue_reset_packed(struct vring_virtqueue *vq)
> >  {
> >         memset(vq->packed.vring.device, 0, vq->packed.event_size_in_bytes);
> > @@ -2172,13 +2539,14 @@ static struct virtqueue *__vring_new_virtqueue_packed(unsigned int index,
> >  #else
> >         vq->broken = false;
> >  #endif
> > -       vq->layout = PACKED;
> >         vq->map = map;
> >         vq->use_map_api = vring_use_map_api(vdev);
> >
> >         vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
> >                 !context;
> >         vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
> > +       vq->layout = virtio_has_feature(vdev, VIRTIO_F_IN_ORDER) ?
> > +                    PACKED_IN_ORDER : PACKED;
> >
> >         if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
> >                 vq->weak_barriers = false;
> > @@ -2288,9 +2656,39 @@ static const struct virtqueue_ops packed_ops = {
> >         .reset = virtqueue_reset_packed,
> >  };
> >
> > +static const struct virtqueue_ops split_in_order_ops = {
> > +       .add = virtqueue_add_split,
> > +       .get = virtqueue_get_buf_ctx_split_in_order,
> > +       .kick_prepare = virtqueue_kick_prepare_split,
> > +       .disable_cb = virtqueue_disable_cb_split,
> > +       .enable_cb_delayed = virtqueue_enable_cb_delayed_split,
> > +       .enable_cb_prepare = virtqueue_enable_cb_prepare_split,
> > +       .poll = virtqueue_poll_split,
> > +       .detach_unused_buf = virtqueue_detach_unused_buf_split,
> > +       .more_used = more_used_split_in_order,
> > +       .resize = virtqueue_resize_split,
> > +       .reset = virtqueue_reset_split,
> > +};
> > +
> > +static const struct virtqueue_ops packed_in_order_ops = {
> > +       .add = virtqueue_add_packed_in_order,
> > +       .get = virtqueue_get_buf_ctx_packed_in_order,
> > +       .kick_prepare = virtqueue_kick_prepare_packed,
> > +       .disable_cb = virtqueue_disable_cb_packed,
> > +       .enable_cb_delayed = virtqueue_enable_cb_delayed_packed,
> > +       .enable_cb_prepare = virtqueue_enable_cb_prepare_packed,
> > +       .poll = virtqueue_poll_packed,
> > +       .detach_unused_buf = virtqueue_detach_unused_buf_packed,
> > +       .more_used = more_used_packed_in_order,
> > +       .resize = virtqueue_resize_packed,
> > +       .reset = virtqueue_reset_packed,
> > +};
> > +
> >  static const struct virtqueue_ops *const all_ops[VQ_TYPE_MAX] = {
> >         [SPLIT] = &split_ops,
> > -       [PACKED] = &packed_ops
> > +       [PACKED] = &packed_ops,
> > +       [SPLIT_IN_ORDER] = &split_in_order_ops,
> > +       [PACKED_IN_ORDER] = &packed_in_order_ops,
> >  };
> >
> >  static int virtqueue_disable_and_recycle(struct virtqueue *_vq,
> > @@ -2346,6 +2744,12 @@ static int virtqueue_enable_after_reset(struct virtqueue *_vq)
> >         case PACKED:                                                    \
> >                 ret = all_ops[PACKED]->op(vq, ##__VA_ARGS__);           \
> >                 break;                                                  \
> > +       case SPLIT_IN_ORDER:                                            \
> > +               ret = all_ops[SPLIT_IN_ORDER]->op(vq, ##__VA_ARGS__);   \
> > +               break;                                                  \
> > +       case PACKED_IN_ORDER:                                           \
> > +               ret = all_ops[PACKED_IN_ORDER]->op(vq, ##__VA_ARGS__);  \
> > +               break;                                                  \
> >         default:                                                        \
> >                 BUG();                                                  \
> >                 break;                                                  \
> > @@ -2362,10 +2766,16 @@ static int virtqueue_enable_after_reset(struct virtqueue *_vq)
> >         case PACKED:                                    \
> >                 all_ops[PACKED]->op(vq, ##__VA_ARGS__); \
> >                 break;                                  \
> > -       default:                                        \
> > -               BUG();                                  \
> > -               break;                                  \
> > -       }                                               \
> > +       case SPLIT_IN_ORDER:                                            \
> > +               all_ops[SPLIT_IN_ORDER]->op(vq, ##__VA_ARGS__); \
> > +               break;                                                  \
> > +       case PACKED_IN_ORDER:                                           \
> > +               all_ops[PACKED_IN_ORDER]->op(vq, ##__VA_ARGS__);        \
> > +               break;                                                  \
> > +       default:                                                        \
> > +               BUG();                                                  \
> > +               break;                                                  \
> > +       }                                                               \
> >  })
> >
> >  static inline int virtqueue_add(struct virtqueue *_vq,
> > @@ -3081,6 +3491,8 @@ void vring_transport_features(struct virtio_device *vdev)
> >                         break;
> >                 case VIRTIO_F_NOTIFICATION_DATA:
> >                         break;
> > +               case VIRTIO_F_IN_ORDER:
> > +                       break;
> >                 default:
> >                         /* We don't understand this bit. */
> >                         __virtio_clear_bit(vdev, i);
> > --
> > 2.31.1
> >

Re: [PATCH V9 19/19] virtio_ring: add in order support

[Jason Wang]

On Fri, Dec 19, 2025 at 7:04 PM Eugenio Perez Martin
<eperezma@redhat.com> wrote:
>
> On Fri, Dec 19, 2025 at 11:44 AM Eugenio Perez Martin
> <eperezma@redhat.com> wrote:
> >
> > On Tue, Nov 25, 2025 at 8:10 AM Jason Wang <jasowang@redhat.com> wrote:
> > >
> > > This patch implements in order support for both split virtqueue and
> > > packed virtqueue. Performance could be gained for the device where the
> > > memory access could be expensive (e.g vhost-net or a real PCI device):
> > >
> > > Benchmark with KVM guest:
> > >
> > > Vhost-net on the host: (pktgen + XDP_DROP):
> > >
> > >          in_order=off | in_order=on | +%
> > >     TX:  4.51Mpps     | 5.30Mpps    | +17%
> > >     RX:  3.47Mpps     | 3.61Mpps    | + 4%
> > >
> > > Vhost-user(testpmd) on the host: (pktgen/XDP_DROP):
> > >
> > > For split virtqueue:
> > >
> > >          in_order=off | in_order=on | +%
> > >     TX:  5.60Mpps     | 5.60Mpps    | +0.0%
> > >     RX:  9.16Mpps     | 9.61Mpps    | +4.9%
> > >
> > > For packed virtqueue:
> > >
> > >          in_order=off | in_order=on | +%
> > >     TX:  5.60Mpps     | 5.70Mpps    | +1.7%
> > >     RX:  10.6Mpps     | 10.8Mpps    | +1.8%
> > >
> > > Benchmark also shows no performance impact for in_order=off for queue
> > > size with 256 and 1024.
> > >
> > > Signed-off-by: Jason Wang <jasowang@redhat.com>
> >
> > Reviewed-by: Eugenio Pérez <eperezma@redhat.com>
> >
> > I have one nitpick though, but it can be added on top or just ignored.
> >
> > > ---
> > >  drivers/virtio/virtio_ring.c | 458 +++++++++++++++++++++++++++++++++--
> > >  1 file changed, 435 insertions(+), 23 deletions(-)
> > >
> > > diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> > > index 61884e031b94..9847b23f82ea 100644
> > > --- a/drivers/virtio/virtio_ring.c
> > > +++ b/drivers/virtio/virtio_ring.c
> > > @@ -70,6 +70,8 @@
> > >  enum vq_layout {
> > >         SPLIT = 0,
> > >         PACKED,
> > > +       SPLIT_IN_ORDER,
> > > +       PACKED_IN_ORDER,
> > >         VQ_TYPE_MAX,
> > >  };
> > >
> > > @@ -80,6 +82,7 @@ struct vring_desc_state_split {
> > >          * allocated together. So we won't stress more to the memory allocator.
> > >          */
> > >         struct vring_desc *indir_desc;
> > > +       u32 total_len;                  /* Buffer Length */
> > >  };
> > >
> > >  struct vring_desc_state_packed {
> > > @@ -91,6 +94,7 @@ struct vring_desc_state_packed {
> > >         struct vring_packed_desc *indir_desc;
> > >         u16 num;                        /* Descriptor list length. */
> > >         u16 last;                       /* The last desc state in a list. */
> > > +       u32 total_len;                  /* Buffer Length */
> > >  };
> > >
> > >  struct vring_desc_extra {
> > > @@ -205,8 +209,24 @@ struct vring_virtqueue {
> > >
> > >         enum vq_layout layout;
> > >
> > > -       /* Head of free buffer list. */
> > > +       /*
> > > +        * Without IN_ORDER it's the head of free buffer list. With
> > > +        * IN_ORDER and SPLIT, it's the next available buffer
> > > +        * index. With IN_ORDER and PACKED, it's unused.
> > > +        */
> > >         unsigned int free_head;
> > > +
> > > +       /*
> > > +        * With IN_ORDER, once we see an in-order batch, this stores
> > > +        * this last entry, and until we return the last buffer.
> > > +        * After this, id is set to vq.num to mark it invalid.
> > > +        * Unused without IN_ORDER.
> > > +        */
> > > +       struct used_entry {
> > > +               u32 id;
> > > +               u32 len;
> > > +       } batch_last;
> > > +
> > >         /* Number we've added since last sync. */
> > >         unsigned int num_added;
> > >
> > > @@ -218,6 +238,11 @@ struct vring_virtqueue {
> > >          */
> > >         u16 last_used_idx;
> > >
> > > +       /* With IN_ORDER and SPLIT, last descriptor id we used to
> > > +        * detach buffer.
> > > +        */
> > > +       u16 last_used;
> > > +
> > >         /* Hint for event idx: already triggered no need to disable. */
> > >         bool event_triggered;
> > >
> > > @@ -259,7 +284,12 @@ static void vring_free(struct virtqueue *_vq);
> > >
> > >  static inline bool virtqueue_is_packed(const struct vring_virtqueue *vq)
> > >  {
> > > -       return vq->layout == PACKED;
> > > +       return vq->layout == PACKED || vq->layout == PACKED_IN_ORDER;
> > > +}
> > > +
> > > +static inline bool virtqueue_is_in_order(const struct vring_virtqueue *vq)
> > > +{
> > > +       return vq->layout == SPLIT_IN_ORDER || vq->layout == PACKED_IN_ORDER;
> > >  }
> > >
> > >  static bool virtqueue_use_indirect(const struct vring_virtqueue *vq,
> > > @@ -469,6 +499,8 @@ static void virtqueue_init(struct vring_virtqueue *vq, u32 num)
> > >         else
> > >                 vq->last_used_idx = 0;
> > >
> > > +       vq->last_used = 0;
> > > +
> > >         vq->event_triggered = false;
> > >         vq->num_added = 0;
> > >
> > > @@ -576,6 +608,8 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
> > >         struct scatterlist *sg;
> > >         struct vring_desc *desc;
> > >         unsigned int i, n, avail, descs_used, err_idx, sg_count = 0;
> > > +       /* Total length for in-order */
> > > +       unsigned int total_len = 0;
> > >         int head;
> > >         bool indirect;
> > >
> > > @@ -648,6 +682,7 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
> > >                          */
> > >                         i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
> > >                                                      len, flags ,premapped);
> > > +                       total_len += len;
> > >                 }
> > >         }
> > >         for (; n < (out_sgs + in_sgs); n++) {
> > > @@ -667,6 +702,7 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
> > >                          */
> > >                         i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
> > >                                                      len, flags, premapped);
> > > +                       total_len += len;
> > >                 }
> > >         }
> > >
> > > @@ -689,7 +725,12 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
> > >         vq->vq.num_free -= descs_used;
> > >
> > >         /* Update free pointer */
> > > -       if (indirect)
> > > +       if (virtqueue_is_in_order(vq)) {
> > > +               vq->free_head += descs_used;
> > > +               if (vq->free_head >= vq->split.vring.num)
> > > +                       vq->free_head -= vq->split.vring.num;
> > > +               vq->split.desc_state[head].total_len = total_len;
> > > +       } else if (indirect)
> > >                 vq->free_head = vq->split.desc_extra[head].next;
> > >         else
> > >                 vq->free_head = i;
> > > @@ -862,6 +903,14 @@ static bool more_used_split(const struct vring_virtqueue *vq)
> > >         return virtqueue_poll_split(vq, vq->last_used_idx);
> > >  }
> > >
> > > +static bool more_used_split_in_order(const struct vring_virtqueue *vq)
> > > +{
> > > +       if (vq->batch_last.id != UINT_MAX)
> > > +               return true;
> > > +
> > > +       return virtqueue_poll_split(vq, vq->last_used_idx);
> > > +}
> > > +
> > >  static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
> > >                                          unsigned int *len,
> > >                                          void **ctx)
> > > @@ -919,6 +968,80 @@ static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
> > >         return ret;
> > >  }
> > >
> > > +static void *virtqueue_get_buf_ctx_split_in_order(struct vring_virtqueue *vq,
> > > +                                                 unsigned int *len,
> > > +                                                 void **ctx)
> > > +{
> > > +       void *ret;
> > > +       unsigned int num = vq->split.vring.num;
> > > +       unsigned int num_free = vq->vq.num_free;
> > > +       u16 last_used, last_used_idx;
> > > +
> > > +       START_USE(vq);
> > > +
> > > +       if (unlikely(vq->broken)) {
> > > +               END_USE(vq);
> > > +               return NULL;
> > > +       }
> > > +
> > > +       last_used = vq->last_used & (num - 1);
> > > +       last_used_idx = vq->last_used_idx & (num - 1);
> > > +
> > > +       if (vq->batch_last.id == UINT_MAX) {
> > > +               if (!more_used_split_in_order(vq)) {
> > > +                       pr_debug("No more buffers in queue\n");
> > > +                       END_USE(vq);
> > > +                       return NULL;
> > > +               }
> > > +
> > > +               /*
> > > +                * Only get used array entries after they have been
> > > +                * exposed by host.
> > > +                */
> > > +               virtio_rmb(vq->weak_barriers);
> > > +
> > > +               vq->batch_last.id = virtio32_to_cpu(vq->vq.vdev,
> > > +                                   vq->split.vring.used->ring[last_used_idx].id);
> > > +               vq->batch_last.len = virtio32_to_cpu(vq->vq.vdev,
> > > +                                    vq->split.vring.used->ring[last_used_idx].len);
> > > +       }
> > > +
> > > +       if (vq->batch_last.id == last_used) {
> > > +               vq->batch_last.id = UINT_MAX;
> > > +               *len = vq->batch_last.len;
> > > +       } else {
> > > +               *len = vq->split.desc_state[last_used].total_len;
> > > +       }
> > > +
> > > +       if (unlikely(last_used >= num)) {
> > > +               BAD_RING(vq, "id %u out of range\n", last_used);
> > > +               return NULL;
> > > +       }
> > > +       if (unlikely(!vq->split.desc_state[last_used].data)) {
> > > +               BAD_RING(vq, "id %u is not a head!\n", last_used);
> > > +               return NULL;
> > > +       }
> > > +
> > > +       /* detach_buf_split clears data, so grab it now. */
> > > +       ret = vq->split.desc_state[last_used].data;
> > > +       detach_buf_split_in_order(vq, last_used, ctx);
> > > +
> > > +       vq->last_used_idx++;
> > > +       vq->last_used += (vq->vq.num_free - num_free);
> > > +       /* If we expect an interrupt for the next entry, tell host
> > > +        * by writing event index and flush out the write before
> > > +        * the read in the next get_buf call. */
> > > +       if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
> > > +               virtio_store_mb(vq->weak_barriers,
> > > +                               &vring_used_event(&vq->split.vring),
> > > +                               cpu_to_virtio16(vq->vq.vdev, vq->last_used_idx));
> > > +
> > > +       LAST_ADD_TIME_INVALID(vq);
> > > +
> > > +       END_USE(vq);
> > > +       return ret;
> > > +}
> > > +
> > >  static void virtqueue_disable_cb_split(struct vring_virtqueue *vq)
> > >  {
> > >         if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
> > > @@ -1012,7 +1135,10 @@ static void *virtqueue_detach_unused_buf_split(struct vring_virtqueue *vq)
> > >                         continue;
> > >                 /* detach_buf_split clears data, so grab it now. */
> > >                 buf = vq->split.desc_state[i].data;
> > > -               detach_buf_split(vq, i, NULL);
> > > +               if (virtqueue_is_in_order(vq))
> > > +                       detach_buf_split_in_order(vq, i, NULL);
> > > +               else
> > > +                       detach_buf_split(vq, i, NULL);
> > >                 vq->split.avail_idx_shadow--;
> > >                 vq->split.vring.avail->idx = cpu_to_virtio16(vq->vq.vdev,
> > >                                 vq->split.avail_idx_shadow);
> > > @@ -1075,6 +1201,7 @@ static void virtqueue_vring_attach_split(struct vring_virtqueue *vq,
> > >
> > >         /* Put everything in free lists. */
> > >         vq->free_head = 0;
> > > +       vq->batch_last.id = UINT_MAX;
> > >  }
> > >
> > >  static int vring_alloc_state_extra_split(struct vring_virtqueue_split *vring_split)
> > > @@ -1186,7 +1313,6 @@ static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
> > >         if (!vq)
> > >                 return NULL;
> > >
> > > -       vq->layout = SPLIT;
> > >         vq->vq.callback = callback;
> > >         vq->vq.vdev = vdev;
> > >         vq->vq.name = name;
> > > @@ -1206,6 +1332,8 @@ static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
> > >         vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
> > >                 !context;
> > >         vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
> > > +       vq->layout = virtio_has_feature(vdev, VIRTIO_F_IN_ORDER) ?
> > > +                    SPLIT_IN_ORDER : SPLIT;
> > >
> > >         if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
> > >                 vq->weak_barriers = false;
> > > @@ -1363,13 +1491,14 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
> > >                                          unsigned int in_sgs,
> > >                                          void *data,
> > >                                          bool premapped,
> > > -                                        gfp_t gfp)
> > > +                                        gfp_t gfp,
> > > +                                        u16 id)
> > >  {
> > >         struct vring_desc_extra *extra;
> > >         struct vring_packed_desc *desc;
> > >         struct scatterlist *sg;
> > > -       unsigned int i, n, err_idx, len;
> > > -       u16 head, id;
> > > +       unsigned int i, n, err_idx, len, total_len = 0;
> > > +       u16 head;
> > >         dma_addr_t addr;
> > >
> > >         head = vq->packed.next_avail_idx;
> > > @@ -1387,8 +1516,6 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
> > >         }
> > >
> > >         i = 0;
> > > -       id = vq->free_head;
> > > -       BUG_ON(id == vq->packed.vring.num);
> > >
> > >         for (n = 0; n < out_sgs + in_sgs; n++) {
> > >                 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
> > > @@ -1408,6 +1535,7 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
> > >                                 extra[i].flags = n < out_sgs ?  0 : VRING_DESC_F_WRITE;
> > >                         }
> > >
> > > +                       total_len += len;
> > >                         i++;
> > >                 }
> > >         }
> > > @@ -1454,13 +1582,15 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
> > >                                 1 << VRING_PACKED_DESC_F_USED;
> > >         }
> > >         vq->packed.next_avail_idx = n;
> > > -       vq->free_head = vq->packed.desc_extra[id].next;
> > > +       if (!virtqueue_is_in_order(vq))
> > > +               vq->free_head = vq->packed.desc_extra[id].next;
> > >
> > >         /* Store token and indirect buffer state. */
> > >         vq->packed.desc_state[id].num = 1;
> > >         vq->packed.desc_state[id].data = data;
> > >         vq->packed.desc_state[id].indir_desc = desc;
> > >         vq->packed.desc_state[id].last = id;
> > > +       vq->packed.desc_state[id].total_len = total_len;
> > >
> > >         vq->num_added += 1;
> > >
> > > @@ -1513,8 +1643,11 @@ static inline int virtqueue_add_packed(struct vring_virtqueue *vq,
> > >         BUG_ON(total_sg == 0);
> > >
> > >         if (virtqueue_use_indirect(vq, total_sg)) {
> > > +               id = vq->free_head;
> > > +               BUG_ON(id == vq->packed.vring.num);
> > >                 err = virtqueue_add_indirect_packed(vq, sgs, total_sg, out_sgs,
> > > -                                                   in_sgs, data, premapped, gfp);
> > > +                                                   in_sgs, data, premapped,
> > > +                                                   gfp, id);
> > >                 if (err != -ENOMEM) {
> > >                         END_USE(vq);
> > >                         return err;
> > > @@ -1635,6 +1768,156 @@ static inline int virtqueue_add_packed(struct vring_virtqueue *vq,
> > >         return -EIO;
> > >  }
> > >
> > > +static inline int virtqueue_add_packed_in_order(struct vring_virtqueue *vq,
> > > +                                               struct scatterlist *sgs[],
> > > +                                               unsigned int total_sg,
> > > +                                               unsigned int out_sgs,
> > > +                                               unsigned int in_sgs,
> > > +                                               void *data,
> > > +                                               void *ctx,
> > > +                                               bool premapped,
> > > +                                               gfp_t gfp)
> > > +{
> > > +       struct vring_packed_desc *desc;
> > > +       struct scatterlist *sg;
> > > +       unsigned int i, n, sg_count, err_idx, total_len = 0;
> > > +       __le16 head_flags, flags;
> > > +       u16 head, avail_used_flags;
> > > +       int err;
> > > +
> > > +       START_USE(vq);
> > > +
> > > +       BUG_ON(data == NULL);
> > > +       BUG_ON(ctx && vq->indirect);
> > > +
> > > +       if (unlikely(vq->broken)) {
> > > +               END_USE(vq);
> > > +               return -EIO;
> > > +       }
> > > +
> > > +       LAST_ADD_TIME_UPDATE(vq);
> > > +
> > > +       BUG_ON(total_sg == 0);
> > > +
> > > +       if (virtqueue_use_indirect(vq, total_sg)) {
> > > +               err = virtqueue_add_indirect_packed(vq, sgs, total_sg, out_sgs,
> > > +                                                   in_sgs, data, premapped, gfp,
> > > +                                                   vq->packed.next_avail_idx);
> > > +               if (err != -ENOMEM) {
> > > +                       END_USE(vq);
> > > +                       return err;
> > > +               }
> > > +
> > > +               /* fall back on direct */
> > > +       }
> > > +
> > > +       head = vq->packed.next_avail_idx;
> > > +       avail_used_flags = vq->packed.avail_used_flags;
> > > +
> > > +       WARN_ON_ONCE(total_sg > vq->packed.vring.num && !vq->indirect);
> > > +
> > > +       desc = vq->packed.vring.desc;
> > > +       i = head;
> > > +
> > > +       if (unlikely(vq->vq.num_free < total_sg)) {
> > > +               pr_debug("Can't add buf len %i - avail = %i\n",
> > > +                        total_sg, vq->vq.num_free);
> > > +               END_USE(vq);
> > > +               return -ENOSPC;
> > > +       }
> > > +
> > > +       sg_count = 0;
> > > +       for (n = 0; n < out_sgs + in_sgs; n++) {
> > > +               for (sg = sgs[n]; sg; sg = sg_next(sg)) {
> > > +                       dma_addr_t addr;
> > > +                       u32 len;
> > > +
> > > +                       flags = 0;
> > > +                       if (++sg_count != total_sg)
> > > +                               flags |= cpu_to_le16(VRING_DESC_F_NEXT);
> > > +                       if (n >= out_sgs)
> > > +                               flags |= cpu_to_le16(VRING_DESC_F_WRITE);
> > > +
> > > +                       if (vring_map_one_sg(vq, sg, n < out_sgs ?
> > > +                                            DMA_TO_DEVICE : DMA_FROM_DEVICE,
> > > +                                            &addr, &len, premapped))
> > > +                               goto unmap_release;
> > > +
> > > +                       flags |= cpu_to_le16(vq->packed.avail_used_flags);
> > > +
> > > +                       if (i == head)
> > > +                               head_flags = flags;
> > > +                       else
> > > +                               desc[i].flags = flags;
> > > +
> > > +                       desc[i].addr = cpu_to_le64(addr);
> > > +                       desc[i].len = cpu_to_le32(len);
> > > +                       desc[i].id = cpu_to_le16(head);
> > > +
> > > +                       if (unlikely(vq->use_map_api)) {
> > > +                               vq->packed.desc_extra[i].addr = premapped ?
> > > +                                     DMA_MAPPING_ERROR: addr;
> > > +                               vq->packed.desc_extra[i].len = len;
> > > +                               vq->packed.desc_extra[i].flags =
> > > +                                       le16_to_cpu(flags);
> > > +                       }
> > > +
> > > +                       if ((unlikely(++i >= vq->packed.vring.num))) {
> > > +                               i = 0;
> > > +                               vq->packed.avail_used_flags ^=
> > > +                                       1 << VRING_PACKED_DESC_F_AVAIL |
> > > +                                       1 << VRING_PACKED_DESC_F_USED;
> > > +                               vq->packed.avail_wrap_counter ^= 1;
> > > +                       }
> > > +
> > > +                       total_len += len;
>
> I hit send too fast :). The used descriptor "length" field is the
> amount of bytes written by the device. So total_len should only be
> updated in the in descs, not in the out ones. Same with split. Maybe
> rename to total_in_len is also useful, I'm calling that way in SVQ.

Yes, I will fix this.

>
> > > +               }
> > > +       }
> > > +
> > > +       /* We're using some buffers from the free list. */
> > > +       vq->vq.num_free -= total_sg;
> > > +
> > > +       /* Update free pointer */
> > > +       vq->packed.next_avail_idx = i;
> > > +
> > > +       /* Store token. */
> > > +       vq->packed.desc_state[head].num = total_sg;
> > > +       vq->packed.desc_state[head].data = data;
> > > +       vq->packed.desc_state[head].indir_desc = ctx;
> > > +       vq->packed.desc_state[head].total_len = total_len;
> > > +
> > > +       /*
> > > +        * A driver MUST NOT make the first descriptor in the list
> > > +        * available before all subsequent descriptors comprising
> > > +        * the list are made available.
> > > +        */
> > > +       virtio_wmb(vq->weak_barriers);
> > > +       vq->packed.vring.desc[head].flags = head_flags;
> > > +       vq->num_added += total_sg;
> > > +
> > > +       pr_debug("Added buffer head %i to %p\n", head, vq);
> > > +       END_USE(vq);
> > > +
> > > +       return 0;
> > > +
> > > +unmap_release:
> > > +       err_idx = i;
> > > +       i = head;
> > > +       vq->packed.avail_used_flags = avail_used_flags;
> > > +
> > > +       for (n = 0; n < total_sg; n++) {
> > > +               if (i == err_idx)
> > > +                       break;
> > > +               vring_unmap_extra_packed(vq, &vq->packed.desc_extra[i]);
> > > +               i++;
> > > +               if (i >= vq->packed.vring.num)
> > > +                       i = 0;
> > > +       }
> > > +
> > > +       END_USE(vq);
> > > +       return -EIO;
> > > +}
> >
> > This function duplicates a lot of logic from virtqueue_add_packed in
> > my opinion. To make descriptors available is the same process in
> > order, except for saving total_len. I like this new version more, as
> > it gets rid of descs_used, that is always = total_sg. Maybe it is
> > better to just perform these steps conditionally, the same way as
> > virtqueue_add_split?
> >

I prefer to leave this as is to reduce the change since it's in the
merge window now.

Thanks

Re: [PATCH V9 19/19] virtio_ring: add in order support

[Eugenio Perez Martin]

On Tue, Nov 25, 2025 at 8:10 AM Jason Wang <jasowang@redhat.com> wrote:
>
> This patch implements in order support for both split virtqueue and
> packed virtqueue. Performance could be gained for the device where the
> memory access could be expensive (e.g vhost-net or a real PCI device):
>
> Benchmark with KVM guest:
>
> Vhost-net on the host: (pktgen + XDP_DROP):
>
>          in_order=off | in_order=on | +%
>     TX:  4.51Mpps     | 5.30Mpps    | +17%
>     RX:  3.47Mpps     | 3.61Mpps    | + 4%
>
> Vhost-user(testpmd) on the host: (pktgen/XDP_DROP):
>
> For split virtqueue:
>
>          in_order=off | in_order=on | +%
>     TX:  5.60Mpps     | 5.60Mpps    | +0.0%
>     RX:  9.16Mpps     | 9.61Mpps    | +4.9%
>
> For packed virtqueue:
>
>          in_order=off | in_order=on | +%
>     TX:  5.60Mpps     | 5.70Mpps    | +1.7%
>     RX:  10.6Mpps     | 10.8Mpps    | +1.8%
>
> Benchmark also shows no performance impact for in_order=off for queue
> size with 256 and 1024.
>
> Signed-off-by: Jason Wang <jasowang@redhat.com>
> ---
>  drivers/virtio/virtio_ring.c | 458 +++++++++++++++++++++++++++++++++--
>  1 file changed, 435 insertions(+), 23 deletions(-)
>
> diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> index 61884e031b94..9847b23f82ea 100644
> --- a/drivers/virtio/virtio_ring.c
> +++ b/drivers/virtio/virtio_ring.c
> @@ -70,6 +70,8 @@
>  enum vq_layout {
>         SPLIT = 0,
>         PACKED,
> +       SPLIT_IN_ORDER,
> +       PACKED_IN_ORDER,
>         VQ_TYPE_MAX,
>  };
>
> @@ -80,6 +82,7 @@ struct vring_desc_state_split {
>          * allocated together. So we won't stress more to the memory allocator.
>          */
>         struct vring_desc *indir_desc;
> +       u32 total_len;                  /* Buffer Length */
>  };
>
>  struct vring_desc_state_packed {
> @@ -91,6 +94,7 @@ struct vring_desc_state_packed {
>         struct vring_packed_desc *indir_desc;
>         u16 num;                        /* Descriptor list length. */
>         u16 last;                       /* The last desc state in a list. */
> +       u32 total_len;                  /* Buffer Length */
>  };
>
>  struct vring_desc_extra {
> @@ -205,8 +209,24 @@ struct vring_virtqueue {
>
>         enum vq_layout layout;
>
> -       /* Head of free buffer list. */
> +       /*
> +        * Without IN_ORDER it's the head of free buffer list. With
> +        * IN_ORDER and SPLIT, it's the next available buffer
> +        * index. With IN_ORDER and PACKED, it's unused.
> +        */
>         unsigned int free_head;
> +
> +       /*
> +        * With IN_ORDER, once we see an in-order batch, this stores
> +        * this last entry, and until we return the last buffer.
> +        * After this, id is set to vq.num to mark it invalid.

s/vq.num/UINT_MAX/.

> +        * Unused without IN_ORDER.
> +        */
> +       struct used_entry {
> +               u32 id;
> +               u32 len;
> +       } batch_last;
> +
>         /* Number we've added since last sync. */
>         unsigned int num_added;
>
> @@ -218,6 +238,11 @@ struct vring_virtqueue {
>          */
>         u16 last_used_idx;
>
> +       /* With IN_ORDER and SPLIT, last descriptor id we used to
> +        * detach buffer.
> +        */
> +       u16 last_used;
> +
>         /* Hint for event idx: already triggered no need to disable. */
>         bool event_triggered;
>
> @@ -259,7 +284,12 @@ static void vring_free(struct virtqueue *_vq);
>
>  static inline bool virtqueue_is_packed(const struct vring_virtqueue *vq)
>  {
> -       return vq->layout == PACKED;
> +       return vq->layout == PACKED || vq->layout == PACKED_IN_ORDER;
> +}
> +
> +static inline bool virtqueue_is_in_order(const struct vring_virtqueue *vq)
> +{
> +       return vq->layout == SPLIT_IN_ORDER || vq->layout == PACKED_IN_ORDER;
>  }
>
>  static bool virtqueue_use_indirect(const struct vring_virtqueue *vq,
> @@ -469,6 +499,8 @@ static void virtqueue_init(struct vring_virtqueue *vq, u32 num)
>         else
>                 vq->last_used_idx = 0;
>
> +       vq->last_used = 0;
> +
>         vq->event_triggered = false;
>         vq->num_added = 0;
>
> @@ -576,6 +608,8 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
>         struct scatterlist *sg;
>         struct vring_desc *desc;
>         unsigned int i, n, avail, descs_used, err_idx, sg_count = 0;
> +       /* Total length for in-order */
> +       unsigned int total_len = 0;
>         int head;
>         bool indirect;
>
> @@ -648,6 +682,7 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
>                          */
>                         i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
>                                                      len, flags ,premapped);
> +                       total_len += len;
>                 }
>         }
>         for (; n < (out_sgs + in_sgs); n++) {
> @@ -667,6 +702,7 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
>                          */
>                         i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
>                                                      len, flags, premapped);
> +                       total_len += len;
>                 }
>         }
>
> @@ -689,7 +725,12 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
>         vq->vq.num_free -= descs_used;
>
>         /* Update free pointer */
> -       if (indirect)
> +       if (virtqueue_is_in_order(vq)) {
> +               vq->free_head += descs_used;
> +               if (vq->free_head >= vq->split.vring.num)
> +                       vq->free_head -= vq->split.vring.num;
> +               vq->split.desc_state[head].total_len = total_len;
> +       } else if (indirect)
>                 vq->free_head = vq->split.desc_extra[head].next;
>         else
>                 vq->free_head = i;
> @@ -862,6 +903,14 @@ static bool more_used_split(const struct vring_virtqueue *vq)
>         return virtqueue_poll_split(vq, vq->last_used_idx);
>  }
>
> +static bool more_used_split_in_order(const struct vring_virtqueue *vq)
> +{
> +       if (vq->batch_last.id != UINT_MAX)
> +               return true;
> +
> +       return virtqueue_poll_split(vq, vq->last_used_idx);
> +}
> +
>  static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
>                                          unsigned int *len,
>                                          void **ctx)
> @@ -919,6 +968,80 @@ static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
>         return ret;
>  }
>
> +static void *virtqueue_get_buf_ctx_split_in_order(struct vring_virtqueue *vq,
> +                                                 unsigned int *len,
> +                                                 void **ctx)
> +{
> +       void *ret;
> +       unsigned int num = vq->split.vring.num;
> +       unsigned int num_free = vq->vq.num_free;
> +       u16 last_used, last_used_idx;
> +
> +       START_USE(vq);
> +
> +       if (unlikely(vq->broken)) {
> +               END_USE(vq);
> +               return NULL;
> +       }
> +
> +       last_used = vq->last_used & (num - 1);
> +       last_used_idx = vq->last_used_idx & (num - 1);
> +
> +       if (vq->batch_last.id == UINT_MAX) {
> +               if (!more_used_split_in_order(vq)) {
> +                       pr_debug("No more buffers in queue\n");
> +                       END_USE(vq);
> +                       return NULL;
> +               }
> +
> +               /*
> +                * Only get used array entries after they have been
> +                * exposed by host.
> +                */
> +               virtio_rmb(vq->weak_barriers);
> +
> +               vq->batch_last.id = virtio32_to_cpu(vq->vq.vdev,
> +                                   vq->split.vring.used->ring[last_used_idx].id);
> +               vq->batch_last.len = virtio32_to_cpu(vq->vq.vdev,
> +                                    vq->split.vring.used->ring[last_used_idx].len);
> +       }
> +
> +       if (vq->batch_last.id == last_used) {
> +               vq->batch_last.id = UINT_MAX;
> +               *len = vq->batch_last.len;
> +       } else {
> +               *len = vq->split.desc_state[last_used].total_len;
> +       }
> +
> +       if (unlikely(last_used >= num)) {
> +               BAD_RING(vq, "id %u out of range\n", last_used);
> +               return NULL;
> +       }
> +       if (unlikely(!vq->split.desc_state[last_used].data)) {
> +               BAD_RING(vq, "id %u is not a head!\n", last_used);
> +               return NULL;
> +       }
> +
> +       /* detach_buf_split clears data, so grab it now. */
> +       ret = vq->split.desc_state[last_used].data;
> +       detach_buf_split_in_order(vq, last_used, ctx);
> +
> +       vq->last_used_idx++;
> +       vq->last_used += (vq->vq.num_free - num_free);
> +       /* If we expect an interrupt for the next entry, tell host
> +        * by writing event index and flush out the write before
> +        * the read in the next get_buf call. */
> +       if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
> +               virtio_store_mb(vq->weak_barriers,
> +                               &vring_used_event(&vq->split.vring),
> +                               cpu_to_virtio16(vq->vq.vdev, vq->last_used_idx));
> +
> +       LAST_ADD_TIME_INVALID(vq);
> +
> +       END_USE(vq);
> +       return ret;
> +}
> +
>  static void virtqueue_disable_cb_split(struct vring_virtqueue *vq)
>  {
>         if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
> @@ -1012,7 +1135,10 @@ static void *virtqueue_detach_unused_buf_split(struct vring_virtqueue *vq)
>                         continue;
>                 /* detach_buf_split clears data, so grab it now. */
>                 buf = vq->split.desc_state[i].data;
> -               detach_buf_split(vq, i, NULL);
> +               if (virtqueue_is_in_order(vq))
> +                       detach_buf_split_in_order(vq, i, NULL);
> +               else
> +                       detach_buf_split(vq, i, NULL);
>                 vq->split.avail_idx_shadow--;
>                 vq->split.vring.avail->idx = cpu_to_virtio16(vq->vq.vdev,
>                                 vq->split.avail_idx_shadow);
> @@ -1075,6 +1201,7 @@ static void virtqueue_vring_attach_split(struct vring_virtqueue *vq,
>
>         /* Put everything in free lists. */
>         vq->free_head = 0;
> +       vq->batch_last.id = UINT_MAX;
>  }
>
>  static int vring_alloc_state_extra_split(struct vring_virtqueue_split *vring_split)
> @@ -1186,7 +1313,6 @@ static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
>         if (!vq)
>                 return NULL;
>
> -       vq->layout = SPLIT;
>         vq->vq.callback = callback;
>         vq->vq.vdev = vdev;
>         vq->vq.name = name;
> @@ -1206,6 +1332,8 @@ static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
>         vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
>                 !context;
>         vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
> +       vq->layout = virtio_has_feature(vdev, VIRTIO_F_IN_ORDER) ?
> +                    SPLIT_IN_ORDER : SPLIT;
>
>         if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
>                 vq->weak_barriers = false;
> @@ -1363,13 +1491,14 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
>                                          unsigned int in_sgs,
>                                          void *data,
>                                          bool premapped,
> -                                        gfp_t gfp)
> +                                        gfp_t gfp,
> +                                        u16 id)
>  {
>         struct vring_desc_extra *extra;
>         struct vring_packed_desc *desc;
>         struct scatterlist *sg;
> -       unsigned int i, n, err_idx, len;
> -       u16 head, id;
> +       unsigned int i, n, err_idx, len, total_len = 0;
> +       u16 head;
>         dma_addr_t addr;
>
>         head = vq->packed.next_avail_idx;
> @@ -1387,8 +1516,6 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
>         }
>
>         i = 0;
> -       id = vq->free_head;
> -       BUG_ON(id == vq->packed.vring.num);
>
>         for (n = 0; n < out_sgs + in_sgs; n++) {
>                 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
> @@ -1408,6 +1535,7 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
>                                 extra[i].flags = n < out_sgs ?  0 : VRING_DESC_F_WRITE;
>                         }
>
> +                       total_len += len;
>                         i++;
>                 }
>         }
> @@ -1454,13 +1582,15 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
>                                 1 << VRING_PACKED_DESC_F_USED;
>         }
>         vq->packed.next_avail_idx = n;
> -       vq->free_head = vq->packed.desc_extra[id].next;
> +       if (!virtqueue_is_in_order(vq))
> +               vq->free_head = vq->packed.desc_extra[id].next;
>
>         /* Store token and indirect buffer state. */
>         vq->packed.desc_state[id].num = 1;
>         vq->packed.desc_state[id].data = data;
>         vq->packed.desc_state[id].indir_desc = desc;
>         vq->packed.desc_state[id].last = id;
> +       vq->packed.desc_state[id].total_len = total_len;
>
>         vq->num_added += 1;
>
> @@ -1513,8 +1643,11 @@ static inline int virtqueue_add_packed(struct vring_virtqueue *vq,
>         BUG_ON(total_sg == 0);
>
>         if (virtqueue_use_indirect(vq, total_sg)) {
> +               id = vq->free_head;
> +               BUG_ON(id == vq->packed.vring.num);
>                 err = virtqueue_add_indirect_packed(vq, sgs, total_sg, out_sgs,
> -                                                   in_sgs, data, premapped, gfp);
> +                                                   in_sgs, data, premapped,
> +                                                   gfp, id);
>                 if (err != -ENOMEM) {
>                         END_USE(vq);
>                         return err;
> @@ -1635,6 +1768,156 @@ static inline int virtqueue_add_packed(struct vring_virtqueue *vq,
>         return -EIO;
>  }
>
> +static inline int virtqueue_add_packed_in_order(struct vring_virtqueue *vq,
> +                                               struct scatterlist *sgs[],
> +                                               unsigned int total_sg,
> +                                               unsigned int out_sgs,
> +                                               unsigned int in_sgs,
> +                                               void *data,
> +                                               void *ctx,
> +                                               bool premapped,
> +                                               gfp_t gfp)
> +{
> +       struct vring_packed_desc *desc;
> +       struct scatterlist *sg;
> +       unsigned int i, n, sg_count, err_idx, total_len = 0;
> +       __le16 head_flags, flags;
> +       u16 head, avail_used_flags;
> +       int err;
> +
> +       START_USE(vq);
> +
> +       BUG_ON(data == NULL);
> +       BUG_ON(ctx && vq->indirect);
> +
> +       if (unlikely(vq->broken)) {
> +               END_USE(vq);
> +               return -EIO;
> +       }
> +
> +       LAST_ADD_TIME_UPDATE(vq);
> +
> +       BUG_ON(total_sg == 0);
> +
> +       if (virtqueue_use_indirect(vq, total_sg)) {
> +               err = virtqueue_add_indirect_packed(vq, sgs, total_sg, out_sgs,
> +                                                   in_sgs, data, premapped, gfp,
> +                                                   vq->packed.next_avail_idx);
> +               if (err != -ENOMEM) {
> +                       END_USE(vq);
> +                       return err;
> +               }
> +
> +               /* fall back on direct */
> +       }
> +
> +       head = vq->packed.next_avail_idx;
> +       avail_used_flags = vq->packed.avail_used_flags;
> +
> +       WARN_ON_ONCE(total_sg > vq->packed.vring.num && !vq->indirect);
> +
> +       desc = vq->packed.vring.desc;
> +       i = head;
> +
> +       if (unlikely(vq->vq.num_free < total_sg)) {
> +               pr_debug("Can't add buf len %i - avail = %i\n",
> +                        total_sg, vq->vq.num_free);
> +               END_USE(vq);
> +               return -ENOSPC;
> +       }
> +
> +       sg_count = 0;
> +       for (n = 0; n < out_sgs + in_sgs; n++) {
> +               for (sg = sgs[n]; sg; sg = sg_next(sg)) {
> +                       dma_addr_t addr;
> +                       u32 len;
> +
> +                       flags = 0;
> +                       if (++sg_count != total_sg)
> +                               flags |= cpu_to_le16(VRING_DESC_F_NEXT);
> +                       if (n >= out_sgs)
> +                               flags |= cpu_to_le16(VRING_DESC_F_WRITE);
> +
> +                       if (vring_map_one_sg(vq, sg, n < out_sgs ?
> +                                            DMA_TO_DEVICE : DMA_FROM_DEVICE,
> +                                            &addr, &len, premapped))
> +                               goto unmap_release;
> +
> +                       flags |= cpu_to_le16(vq->packed.avail_used_flags);
> +
> +                       if (i == head)
> +                               head_flags = flags;
> +                       else
> +                               desc[i].flags = flags;
> +
> +                       desc[i].addr = cpu_to_le64(addr);
> +                       desc[i].len = cpu_to_le32(len);
> +                       desc[i].id = cpu_to_le16(head);
> +
> +                       if (unlikely(vq->use_map_api)) {
> +                               vq->packed.desc_extra[i].addr = premapped ?
> +                                     DMA_MAPPING_ERROR: addr;
> +                               vq->packed.desc_extra[i].len = len;
> +                               vq->packed.desc_extra[i].flags =
> +                                       le16_to_cpu(flags);
> +                       }
> +
> +                       if ((unlikely(++i >= vq->packed.vring.num))) {
> +                               i = 0;
> +                               vq->packed.avail_used_flags ^=
> +                                       1 << VRING_PACKED_DESC_F_AVAIL |
> +                                       1 << VRING_PACKED_DESC_F_USED;
> +                               vq->packed.avail_wrap_counter ^= 1;
> +                       }
> +
> +                       total_len += len;
> +               }
> +       }
> +
> +       /* We're using some buffers from the free list. */
> +       vq->vq.num_free -= total_sg;
> +
> +       /* Update free pointer */
> +       vq->packed.next_avail_idx = i;
> +
> +       /* Store token. */
> +       vq->packed.desc_state[head].num = total_sg;
> +       vq->packed.desc_state[head].data = data;
> +       vq->packed.desc_state[head].indir_desc = ctx;
> +       vq->packed.desc_state[head].total_len = total_len;
> +
> +       /*
> +        * A driver MUST NOT make the first descriptor in the list
> +        * available before all subsequent descriptors comprising
> +        * the list are made available.
> +        */
> +       virtio_wmb(vq->weak_barriers);
> +       vq->packed.vring.desc[head].flags = head_flags;
> +       vq->num_added += total_sg;
> +
> +       pr_debug("Added buffer head %i to %p\n", head, vq);
> +       END_USE(vq);
> +
> +       return 0;
> +
> +unmap_release:
> +       err_idx = i;
> +       i = head;
> +       vq->packed.avail_used_flags = avail_used_flags;
> +
> +       for (n = 0; n < total_sg; n++) {
> +               if (i == err_idx)
> +                       break;
> +               vring_unmap_extra_packed(vq, &vq->packed.desc_extra[i]);
> +               i++;
> +               if (i >= vq->packed.vring.num)
> +                       i = 0;
> +       }
> +
> +       END_USE(vq);
> +       return -EIO;
> +}
> +
>  static bool virtqueue_kick_prepare_packed(struct vring_virtqueue *vq)
>  {
>         u16 new, old, off_wrap, flags, wrap_counter, event_idx;
> @@ -1796,10 +2079,82 @@ static void update_last_used_idx_packed(struct vring_virtqueue *vq,
>                                 cpu_to_le16(vq->last_used_idx));
>  }
>
> +static bool more_used_packed_in_order(const struct vring_virtqueue *vq)
> +{
> +       if (vq->batch_last.id != UINT_MAX)
> +               return true;
> +
> +       return virtqueue_poll_packed(vq, READ_ONCE(vq->last_used_idx));
> +}
> +
> +static void *virtqueue_get_buf_ctx_packed_in_order(struct vring_virtqueue *vq,
> +                                                  unsigned int *len,
> +                                                  void **ctx)
> +{
> +       unsigned int num = vq->packed.vring.num;
> +       u16 last_used, last_used_idx;
> +       bool used_wrap_counter;
> +       void *ret;
> +
> +       START_USE(vq);
> +
> +       if (unlikely(vq->broken)) {
> +               END_USE(vq);
> +               return NULL;
> +       }
> +
> +       last_used_idx = vq->last_used_idx;
> +       used_wrap_counter = packed_used_wrap_counter(last_used_idx);
> +       last_used = packed_last_used(last_used_idx);
> +
> +       if (vq->batch_last.id == UINT_MAX) {
> +               if (!more_used_packed_in_order(vq)) {
> +                       pr_debug("No more buffers in queue\n");
> +                       END_USE(vq);
> +                       return NULL;
> +               }
> +               /* Only get used elements after they have been exposed by host. */
> +               virtio_rmb(vq->weak_barriers);
> +               vq->batch_last.id =
> +                       le16_to_cpu(vq->packed.vring.desc[last_used].id);
> +               vq->batch_last.len =
> +                       le32_to_cpu(vq->packed.vring.desc[last_used].len);
> +       }
> +
> +       if (vq->batch_last.id == last_used) {
> +               vq->batch_last.id = UINT_MAX;
> +               *len = vq->batch_last.len;
> +       } else {
> +               *len = vq->packed.desc_state[last_used].total_len;
> +       }
> +
> +       if (unlikely(last_used >= num)) {
> +               BAD_RING(vq, "id %u out of range\n", last_used);
> +               return NULL;
> +       }
> +       if (unlikely(!vq->packed.desc_state[last_used].data)) {
> +               BAD_RING(vq, "id %u is not a head!\n", last_used);
> +               return NULL;
> +       }
> +
> +       /* detach_buf_packed clears data, so grab it now. */
> +       ret = vq->packed.desc_state[last_used].data;
> +       detach_buf_packed_in_order(vq, last_used, ctx);
> +
> +       update_last_used_idx_packed(vq, last_used, last_used,
> +                                   used_wrap_counter);
> +
> +       LAST_ADD_TIME_INVALID(vq);
> +
> +       END_USE(vq);
> +       return ret;
> +}
> +
>  static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
>                                           unsigned int *len,
>                                           void **ctx)
>  {
> +       unsigned int num = vq->packed.vring.num;
>         u16 last_used, id, last_used_idx;
>         bool used_wrap_counter;
>         void *ret;
> @@ -1826,7 +2181,7 @@ static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
>         id = le16_to_cpu(vq->packed.vring.desc[last_used].id);
>         *len = le32_to_cpu(vq->packed.vring.desc[last_used].len);
>
> -       if (unlikely(id >= vq->packed.vring.num)) {
> +       if (unlikely(id >= num)) {
>                 BAD_RING(vq, "id %u out of range\n", id);
>                 return NULL;
>         }
> @@ -1967,7 +2322,10 @@ static void *virtqueue_detach_unused_buf_packed(struct vring_virtqueue *vq)
>                         continue;
>                 /* detach_buf clears data, so grab it now. */
>                 buf = vq->packed.desc_state[i].data;
> -               detach_buf_packed(vq, i, NULL);
> +               if (virtqueue_is_in_order(vq))
> +                       detach_buf_packed_in_order(vq, i, NULL);
> +               else
> +                       detach_buf_packed(vq, i, NULL);
>                 END_USE(vq);
>                 return buf;
>         }
> @@ -1993,6 +2351,8 @@ static struct vring_desc_extra *vring_alloc_desc_extra(unsigned int num)
>         for (i = 0; i < num - 1; i++)
>                 desc_extra[i].next = i + 1;
>
> +       desc_extra[num - 1].next = 0;
> +
>         return desc_extra;
>  }
>
> @@ -2124,10 +2484,17 @@ static void virtqueue_vring_attach_packed(struct vring_virtqueue *vq,
>  {
>         vq->packed = *vring_packed;
>
> -       /* Put everything in free lists. */
> -       vq->free_head = 0;
> +       if (virtqueue_is_in_order(vq)) {
> +               vq->batch_last.id = UINT_MAX;
> +       } else {
> +               /*
> +                * Put everything in free lists. Note that
> +                * next_avail_idx is sufficient with IN_ORDER so
> +                * free_head is unused.
> +                */
> +               vq->free_head = 0;
> +       }
>  }
> -
>  static void virtqueue_reset_packed(struct vring_virtqueue *vq)
>  {
>         memset(vq->packed.vring.device, 0, vq->packed.event_size_in_bytes);
> @@ -2172,13 +2539,14 @@ static struct virtqueue *__vring_new_virtqueue_packed(unsigned int index,
>  #else
>         vq->broken = false;
>  #endif
> -       vq->layout = PACKED;
>         vq->map = map;
>         vq->use_map_api = vring_use_map_api(vdev);
>
>         vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
>                 !context;
>         vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
> +       vq->layout = virtio_has_feature(vdev, VIRTIO_F_IN_ORDER) ?
> +                    PACKED_IN_ORDER : PACKED;
>
>         if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
>                 vq->weak_barriers = false;
> @@ -2288,9 +2656,39 @@ static const struct virtqueue_ops packed_ops = {
>         .reset = virtqueue_reset_packed,
>  };
>
> +static const struct virtqueue_ops split_in_order_ops = {
> +       .add = virtqueue_add_split,
> +       .get = virtqueue_get_buf_ctx_split_in_order,
> +       .kick_prepare = virtqueue_kick_prepare_split,
> +       .disable_cb = virtqueue_disable_cb_split,
> +       .enable_cb_delayed = virtqueue_enable_cb_delayed_split,
> +       .enable_cb_prepare = virtqueue_enable_cb_prepare_split,
> +       .poll = virtqueue_poll_split,
> +       .detach_unused_buf = virtqueue_detach_unused_buf_split,
> +       .more_used = more_used_split_in_order,
> +       .resize = virtqueue_resize_split,
> +       .reset = virtqueue_reset_split,
> +};
> +
> +static const struct virtqueue_ops packed_in_order_ops = {
> +       .add = virtqueue_add_packed_in_order,
> +       .get = virtqueue_get_buf_ctx_packed_in_order,
> +       .kick_prepare = virtqueue_kick_prepare_packed,
> +       .disable_cb = virtqueue_disable_cb_packed,
> +       .enable_cb_delayed = virtqueue_enable_cb_delayed_packed,
> +       .enable_cb_prepare = virtqueue_enable_cb_prepare_packed,
> +       .poll = virtqueue_poll_packed,
> +       .detach_unused_buf = virtqueue_detach_unused_buf_packed,
> +       .more_used = more_used_packed_in_order,
> +       .resize = virtqueue_resize_packed,
> +       .reset = virtqueue_reset_packed,
> +};
> +
>  static const struct virtqueue_ops *const all_ops[VQ_TYPE_MAX] = {
>         [SPLIT] = &split_ops,
> -       [PACKED] = &packed_ops
> +       [PACKED] = &packed_ops,
> +       [SPLIT_IN_ORDER] = &split_in_order_ops,
> +       [PACKED_IN_ORDER] = &packed_in_order_ops,
>  };
>
>  static int virtqueue_disable_and_recycle(struct virtqueue *_vq,
> @@ -2346,6 +2744,12 @@ static int virtqueue_enable_after_reset(struct virtqueue *_vq)
>         case PACKED:                                                    \
>                 ret = all_ops[PACKED]->op(vq, ##__VA_ARGS__);           \
>                 break;                                                  \
> +       case SPLIT_IN_ORDER:                                            \
> +               ret = all_ops[SPLIT_IN_ORDER]->op(vq, ##__VA_ARGS__);   \
> +               break;                                                  \
> +       case PACKED_IN_ORDER:                                           \
> +               ret = all_ops[PACKED_IN_ORDER]->op(vq, ##__VA_ARGS__);  \
> +               break;                                                  \
>         default:                                                        \
>                 BUG();                                                  \
>                 break;                                                  \
> @@ -2362,10 +2766,16 @@ static int virtqueue_enable_after_reset(struct virtqueue *_vq)
>         case PACKED:                                    \
>                 all_ops[PACKED]->op(vq, ##__VA_ARGS__); \
>                 break;                                  \
> -       default:                                        \
> -               BUG();                                  \
> -               break;                                  \
> -       }                                               \
> +       case SPLIT_IN_ORDER:                                            \
> +               all_ops[SPLIT_IN_ORDER]->op(vq, ##__VA_ARGS__); \
> +               break;                                                  \
> +       case PACKED_IN_ORDER:                                           \
> +               all_ops[PACKED_IN_ORDER]->op(vq, ##__VA_ARGS__);        \
> +               break;                                                  \
> +       default:                                                        \
> +               BUG();                                                  \
> +               break;                                                  \
> +       }                                                               \
>  })
>
>  static inline int virtqueue_add(struct virtqueue *_vq,
> @@ -3081,6 +3491,8 @@ void vring_transport_features(struct virtio_device *vdev)
>                         break;
>                 case VIRTIO_F_NOTIFICATION_DATA:
>                         break;
> +               case VIRTIO_F_IN_ORDER:
> +                       break;
>                 default:
>                         /* We don't understand this bit. */
>                         __virtio_clear_bit(vdev, i);
> --
> 2.31.1
>

Re: [PATCH V9 19/19] virtio_ring: add in order support

[Eugenio Perez Martin]

On Tue, Nov 25, 2025 at 8:10 AM Jason Wang <jasowang@redhat.com> wrote:
>
> This patch implements in order support for both split virtqueue and
> packed virtqueue. Performance could be gained for the device where the
> memory access could be expensive (e.g vhost-net or a real PCI device):
>
> Benchmark with KVM guest:
>
> Vhost-net on the host: (pktgen + XDP_DROP):
>
>          in_order=off | in_order=on | +%
>     TX:  4.51Mpps     | 5.30Mpps    | +17%
>     RX:  3.47Mpps     | 3.61Mpps    | + 4%
>
> Vhost-user(testpmd) on the host: (pktgen/XDP_DROP):
>
> For split virtqueue:
>
>          in_order=off | in_order=on | +%
>     TX:  5.60Mpps     | 5.60Mpps    | +0.0%
>     RX:  9.16Mpps     | 9.61Mpps    | +4.9%
>
> For packed virtqueue:
>
>          in_order=off | in_order=on | +%
>     TX:  5.60Mpps     | 5.70Mpps    | +1.7%
>     RX:  10.6Mpps     | 10.8Mpps    | +1.8%
>
> Benchmark also shows no performance impact for in_order=off for queue
> size with 256 and 1024.
>
> Signed-off-by: Jason Wang <jasowang@redhat.com>
> ---
>  drivers/virtio/virtio_ring.c | 458 +++++++++++++++++++++++++++++++++--
>  1 file changed, 435 insertions(+), 23 deletions(-)
>
> diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> index 61884e031b94..9847b23f82ea 100644
> --- a/drivers/virtio/virtio_ring.c
> +++ b/drivers/virtio/virtio_ring.c
> @@ -70,6 +70,8 @@
>  enum vq_layout {
>         SPLIT = 0,
>         PACKED,
> +       SPLIT_IN_ORDER,
> +       PACKED_IN_ORDER,
>         VQ_TYPE_MAX,
>  };
>
> @@ -80,6 +82,7 @@ struct vring_desc_state_split {
>          * allocated together. So we won't stress more to the memory allocator.
>          */
>         struct vring_desc *indir_desc;
> +       u32 total_len;                  /* Buffer Length */
>  };
>
>  struct vring_desc_state_packed {
> @@ -91,6 +94,7 @@ struct vring_desc_state_packed {
>         struct vring_packed_desc *indir_desc;
>         u16 num;                        /* Descriptor list length. */
>         u16 last;                       /* The last desc state in a list. */
> +       u32 total_len;                  /* Buffer Length */
>  };
>
>  struct vring_desc_extra {
> @@ -205,8 +209,24 @@ struct vring_virtqueue {
>
>         enum vq_layout layout;
>
> -       /* Head of free buffer list. */
> +       /*
> +        * Without IN_ORDER it's the head of free buffer list. With
> +        * IN_ORDER and SPLIT, it's the next available buffer
> +        * index. With IN_ORDER and PACKED, it's unused.
> +        */
>         unsigned int free_head;
> +
> +       /*
> +        * With IN_ORDER, once we see an in-order batch, this stores
> +        * this last entry, and until we return the last buffer.
> +        * After this, id is set to vq.num to mark it invalid.
> +        * Unused without IN_ORDER.
> +        */
> +       struct used_entry {
> +               u32 id;
> +               u32 len;
> +       } batch_last;
> +
>         /* Number we've added since last sync. */
>         unsigned int num_added;
>
> @@ -218,6 +238,11 @@ struct vring_virtqueue {
>          */
>         u16 last_used_idx;
>
> +       /* With IN_ORDER and SPLIT, last descriptor id we used to
> +        * detach buffer.
> +        */
> +       u16 last_used;
> +
>         /* Hint for event idx: already triggered no need to disable. */
>         bool event_triggered;
>
> @@ -259,7 +284,12 @@ static void vring_free(struct virtqueue *_vq);
>
>  static inline bool virtqueue_is_packed(const struct vring_virtqueue *vq)
>  {
> -       return vq->layout == PACKED;
> +       return vq->layout == PACKED || vq->layout == PACKED_IN_ORDER;
> +}
> +
> +static inline bool virtqueue_is_in_order(const struct vring_virtqueue *vq)
> +{
> +       return vq->layout == SPLIT_IN_ORDER || vq->layout == PACKED_IN_ORDER;
>  }
>
>  static bool virtqueue_use_indirect(const struct vring_virtqueue *vq,
> @@ -469,6 +499,8 @@ static void virtqueue_init(struct vring_virtqueue *vq, u32 num)
>         else
>                 vq->last_used_idx = 0;
>
> +       vq->last_used = 0;
> +
>         vq->event_triggered = false;
>         vq->num_added = 0;
>
> @@ -576,6 +608,8 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
>         struct scatterlist *sg;
>         struct vring_desc *desc;
>         unsigned int i, n, avail, descs_used, err_idx, sg_count = 0;
> +       /* Total length for in-order */
> +       unsigned int total_len = 0;
>         int head;
>         bool indirect;
>
> @@ -648,6 +682,7 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
>                          */
>                         i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
>                                                      len, flags ,premapped);
> +                       total_len += len;
>                 }
>         }
>         for (; n < (out_sgs + in_sgs); n++) {
> @@ -667,6 +702,7 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
>                          */
>                         i = virtqueue_add_desc_split(vq, desc, extra, i, addr,
>                                                      len, flags, premapped);
> +                       total_len += len;
>                 }
>         }
>
> @@ -689,7 +725,12 @@ static inline int virtqueue_add_split(struct vring_virtqueue *vq,
>         vq->vq.num_free -= descs_used;
>
>         /* Update free pointer */
> -       if (indirect)
> +       if (virtqueue_is_in_order(vq)) {
> +               vq->free_head += descs_used;
> +               if (vq->free_head >= vq->split.vring.num)
> +                       vq->free_head -= vq->split.vring.num;
> +               vq->split.desc_state[head].total_len = total_len;
> +       } else if (indirect)
>                 vq->free_head = vq->split.desc_extra[head].next;
>         else
>                 vq->free_head = i;
> @@ -862,6 +903,14 @@ static bool more_used_split(const struct vring_virtqueue *vq)
>         return virtqueue_poll_split(vq, vq->last_used_idx);
>  }
>
> +static bool more_used_split_in_order(const struct vring_virtqueue *vq)
> +{
> +       if (vq->batch_last.id != UINT_MAX)
> +               return true;
> +
> +       return virtqueue_poll_split(vq, vq->last_used_idx);
> +}
> +
>  static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
>                                          unsigned int *len,
>                                          void **ctx)
> @@ -919,6 +968,80 @@ static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
>         return ret;
>  }
>
> +static void *virtqueue_get_buf_ctx_split_in_order(struct vring_virtqueue *vq,
> +                                                 unsigned int *len,
> +                                                 void **ctx)
> +{
> +       void *ret;
> +       unsigned int num = vq->split.vring.num;
> +       unsigned int num_free = vq->vq.num_free;
> +       u16 last_used, last_used_idx;
> +
> +       START_USE(vq);
> +
> +       if (unlikely(vq->broken)) {
> +               END_USE(vq);
> +               return NULL;
> +       }
> +
> +       last_used = vq->last_used & (num - 1);
> +       last_used_idx = vq->last_used_idx & (num - 1);
> +
> +       if (vq->batch_last.id == UINT_MAX) {
> +               if (!more_used_split_in_order(vq)) {
> +                       pr_debug("No more buffers in queue\n");
> +                       END_USE(vq);
> +                       return NULL;
> +               }
> +
> +               /*
> +                * Only get used array entries after they have been
> +                * exposed by host.
> +                */
> +               virtio_rmb(vq->weak_barriers);
> +
> +               vq->batch_last.id = virtio32_to_cpu(vq->vq.vdev,
> +                                   vq->split.vring.used->ring[last_used_idx].id);
> +               vq->batch_last.len = virtio32_to_cpu(vq->vq.vdev,
> +                                    vq->split.vring.used->ring[last_used_idx].len);
> +       }
> +
> +       if (vq->batch_last.id == last_used) {
> +               vq->batch_last.id = UINT_MAX;
> +               *len = vq->batch_last.len;
> +       } else {
> +               *len = vq->split.desc_state[last_used].total_len;
> +       }
> +
> +       if (unlikely(last_used >= num)) {
> +               BAD_RING(vq, "id %u out of range\n", last_used);
> +               return NULL;
> +       }

This cannot be true, as it is ceiled by & (num - 1). Also, this check
should be moved before the access if you remove the ceiling.

> +       if (unlikely(!vq->split.desc_state[last_used].data)) {
> +               BAD_RING(vq, "id %u is not a head!\n", last_used);
> +               return NULL;
> +       }
> +
> +       /* detach_buf_split clears data, so grab it now. */
> +       ret = vq->split.desc_state[last_used].data;
> +       detach_buf_split_in_order(vq, last_used, ctx);
> +
> +       vq->last_used_idx++;
> +       vq->last_used += (vq->vq.num_free - num_free);
> +       /* If we expect an interrupt for the next entry, tell host
> +        * by writing event index and flush out the write before
> +        * the read in the next get_buf call. */
> +       if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
> +               virtio_store_mb(vq->weak_barriers,
> +                               &vring_used_event(&vq->split.vring),
> +                               cpu_to_virtio16(vq->vq.vdev, vq->last_used_idx));
> +
> +       LAST_ADD_TIME_INVALID(vq);
> +
> +       END_USE(vq);
> +       return ret;
> +}
> +
>  static void virtqueue_disable_cb_split(struct vring_virtqueue *vq)
>  {
>         if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
> @@ -1012,7 +1135,10 @@ static void *virtqueue_detach_unused_buf_split(struct vring_virtqueue *vq)
>                         continue;
>                 /* detach_buf_split clears data, so grab it now. */
>                 buf = vq->split.desc_state[i].data;
> -               detach_buf_split(vq, i, NULL);
> +               if (virtqueue_is_in_order(vq))
> +                       detach_buf_split_in_order(vq, i, NULL);
> +               else
> +                       detach_buf_split(vq, i, NULL);
>                 vq->split.avail_idx_shadow--;
>                 vq->split.vring.avail->idx = cpu_to_virtio16(vq->vq.vdev,
>                                 vq->split.avail_idx_shadow);
> @@ -1075,6 +1201,7 @@ static void virtqueue_vring_attach_split(struct vring_virtqueue *vq,
>
>         /* Put everything in free lists. */
>         vq->free_head = 0;
> +       vq->batch_last.id = UINT_MAX;
>  }
>
>  static int vring_alloc_state_extra_split(struct vring_virtqueue_split *vring_split)
> @@ -1186,7 +1313,6 @@ static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
>         if (!vq)
>                 return NULL;
>
> -       vq->layout = SPLIT;
>         vq->vq.callback = callback;
>         vq->vq.vdev = vdev;
>         vq->vq.name = name;
> @@ -1206,6 +1332,8 @@ static struct virtqueue *__vring_new_virtqueue_split(unsigned int index,
>         vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
>                 !context;
>         vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
> +       vq->layout = virtio_has_feature(vdev, VIRTIO_F_IN_ORDER) ?
> +                    SPLIT_IN_ORDER : SPLIT;
>
>         if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
>                 vq->weak_barriers = false;
> @@ -1363,13 +1491,14 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
>                                          unsigned int in_sgs,
>                                          void *data,
>                                          bool premapped,
> -                                        gfp_t gfp)
> +                                        gfp_t gfp,
> +                                        u16 id)
>  {
>         struct vring_desc_extra *extra;
>         struct vring_packed_desc *desc;
>         struct scatterlist *sg;
> -       unsigned int i, n, err_idx, len;
> -       u16 head, id;
> +       unsigned int i, n, err_idx, len, total_len = 0;
> +       u16 head;
>         dma_addr_t addr;
>
>         head = vq->packed.next_avail_idx;
> @@ -1387,8 +1516,6 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
>         }
>
>         i = 0;
> -       id = vq->free_head;
> -       BUG_ON(id == vq->packed.vring.num);
>
>         for (n = 0; n < out_sgs + in_sgs; n++) {
>                 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
> @@ -1408,6 +1535,7 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
>                                 extra[i].flags = n < out_sgs ?  0 : VRING_DESC_F_WRITE;
>                         }
>
> +                       total_len += len;
>                         i++;
>                 }
>         }
> @@ -1454,13 +1582,15 @@ static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
>                                 1 << VRING_PACKED_DESC_F_USED;
>         }
>         vq->packed.next_avail_idx = n;
> -       vq->free_head = vq->packed.desc_extra[id].next;
> +       if (!virtqueue_is_in_order(vq))
> +               vq->free_head = vq->packed.desc_extra[id].next;
>
>         /* Store token and indirect buffer state. */
>         vq->packed.desc_state[id].num = 1;
>         vq->packed.desc_state[id].data = data;
>         vq->packed.desc_state[id].indir_desc = desc;
>         vq->packed.desc_state[id].last = id;
> +       vq->packed.desc_state[id].total_len = total_len;
>
>         vq->num_added += 1;
>
> @@ -1513,8 +1643,11 @@ static inline int virtqueue_add_packed(struct vring_virtqueue *vq,
>         BUG_ON(total_sg == 0);
>
>         if (virtqueue_use_indirect(vq, total_sg)) {
> +               id = vq->free_head;
> +               BUG_ON(id == vq->packed.vring.num);
>                 err = virtqueue_add_indirect_packed(vq, sgs, total_sg, out_sgs,
> -                                                   in_sgs, data, premapped, gfp);
> +                                                   in_sgs, data, premapped,
> +                                                   gfp, id);
>                 if (err != -ENOMEM) {
>                         END_USE(vq);
>                         return err;
> @@ -1635,6 +1768,156 @@ static inline int virtqueue_add_packed(struct vring_virtqueue *vq,
>         return -EIO;
>  }
>
> +static inline int virtqueue_add_packed_in_order(struct vring_virtqueue *vq,
> +                                               struct scatterlist *sgs[],
> +                                               unsigned int total_sg,
> +                                               unsigned int out_sgs,
> +                                               unsigned int in_sgs,
> +                                               void *data,
> +                                               void *ctx,
> +                                               bool premapped,
> +                                               gfp_t gfp)
> +{
> +       struct vring_packed_desc *desc;
> +       struct scatterlist *sg;
> +       unsigned int i, n, sg_count, err_idx, total_len = 0;
> +       __le16 head_flags, flags;
> +       u16 head, avail_used_flags;
> +       int err;
> +
> +       START_USE(vq);
> +
> +       BUG_ON(data == NULL);
> +       BUG_ON(ctx && vq->indirect);
> +
> +       if (unlikely(vq->broken)) {
> +               END_USE(vq);
> +               return -EIO;
> +       }
> +
> +       LAST_ADD_TIME_UPDATE(vq);
> +
> +       BUG_ON(total_sg == 0);
> +
> +       if (virtqueue_use_indirect(vq, total_sg)) {
> +               err = virtqueue_add_indirect_packed(vq, sgs, total_sg, out_sgs,
> +                                                   in_sgs, data, premapped, gfp,
> +                                                   vq->packed.next_avail_idx);
> +               if (err != -ENOMEM) {
> +                       END_USE(vq);
> +                       return err;
> +               }
> +
> +               /* fall back on direct */
> +       }
> +
> +       head = vq->packed.next_avail_idx;
> +       avail_used_flags = vq->packed.avail_used_flags;
> +
> +       WARN_ON_ONCE(total_sg > vq->packed.vring.num && !vq->indirect);
> +
> +       desc = vq->packed.vring.desc;
> +       i = head;
> +
> +       if (unlikely(vq->vq.num_free < total_sg)) {
> +               pr_debug("Can't add buf len %i - avail = %i\n",
> +                        total_sg, vq->vq.num_free);
> +               END_USE(vq);
> +               return -ENOSPC;
> +       }
> +
> +       sg_count = 0;
> +       for (n = 0; n < out_sgs + in_sgs; n++) {
> +               for (sg = sgs[n]; sg; sg = sg_next(sg)) {
> +                       dma_addr_t addr;
> +                       u32 len;
> +
> +                       flags = 0;
> +                       if (++sg_count != total_sg)
> +                               flags |= cpu_to_le16(VRING_DESC_F_NEXT);
> +                       if (n >= out_sgs)
> +                               flags |= cpu_to_le16(VRING_DESC_F_WRITE);
> +
> +                       if (vring_map_one_sg(vq, sg, n < out_sgs ?
> +                                            DMA_TO_DEVICE : DMA_FROM_DEVICE,
> +                                            &addr, &len, premapped))
> +                               goto unmap_release;
> +
> +                       flags |= cpu_to_le16(vq->packed.avail_used_flags);
> +
> +                       if (i == head)
> +                               head_flags = flags;
> +                       else
> +                               desc[i].flags = flags;
> +
> +                       desc[i].addr = cpu_to_le64(addr);
> +                       desc[i].len = cpu_to_le32(len);
> +                       desc[i].id = cpu_to_le16(head);
> +
> +                       if (unlikely(vq->use_map_api)) {
> +                               vq->packed.desc_extra[i].addr = premapped ?
> +                                     DMA_MAPPING_ERROR: addr;
> +                               vq->packed.desc_extra[i].len = len;
> +                               vq->packed.desc_extra[i].flags =
> +                                       le16_to_cpu(flags);
> +                       }
> +
> +                       if ((unlikely(++i >= vq->packed.vring.num))) {
> +                               i = 0;
> +                               vq->packed.avail_used_flags ^=
> +                                       1 << VRING_PACKED_DESC_F_AVAIL |
> +                                       1 << VRING_PACKED_DESC_F_USED;
> +                               vq->packed.avail_wrap_counter ^= 1;
> +                       }
> +
> +                       total_len += len;
> +               }
> +       }
> +
> +       /* We're using some buffers from the free list. */
> +       vq->vq.num_free -= total_sg;
> +
> +       /* Update free pointer */
> +       vq->packed.next_avail_idx = i;
> +
> +       /* Store token. */
> +       vq->packed.desc_state[head].num = total_sg;
> +       vq->packed.desc_state[head].data = data;
> +       vq->packed.desc_state[head].indir_desc = ctx;
> +       vq->packed.desc_state[head].total_len = total_len;
> +
> +       /*
> +        * A driver MUST NOT make the first descriptor in the list
> +        * available before all subsequent descriptors comprising
> +        * the list are made available.
> +        */
> +       virtio_wmb(vq->weak_barriers);
> +       vq->packed.vring.desc[head].flags = head_flags;
> +       vq->num_added += total_sg;
> +
> +       pr_debug("Added buffer head %i to %p\n", head, vq);
> +       END_USE(vq);
> +
> +       return 0;
> +
> +unmap_release:
> +       err_idx = i;
> +       i = head;
> +       vq->packed.avail_used_flags = avail_used_flags;
> +
> +       for (n = 0; n < total_sg; n++) {
> +               if (i == err_idx)
> +                       break;
> +               vring_unmap_extra_packed(vq, &vq->packed.desc_extra[i]);
> +               i++;
> +               if (i >= vq->packed.vring.num)
> +                       i = 0;
> +       }
> +
> +       END_USE(vq);
> +       return -EIO;
> +}
> +
>  static bool virtqueue_kick_prepare_packed(struct vring_virtqueue *vq)
>  {
>         u16 new, old, off_wrap, flags, wrap_counter, event_idx;
> @@ -1796,10 +2079,82 @@ static void update_last_used_idx_packed(struct vring_virtqueue *vq,
>                                 cpu_to_le16(vq->last_used_idx));
>  }
>
> +static bool more_used_packed_in_order(const struct vring_virtqueue *vq)
> +{
> +       if (vq->batch_last.id != UINT_MAX)
> +               return true;
> +
> +       return virtqueue_poll_packed(vq, READ_ONCE(vq->last_used_idx));
> +}
> +
> +static void *virtqueue_get_buf_ctx_packed_in_order(struct vring_virtqueue *vq,
> +                                                  unsigned int *len,
> +                                                  void **ctx)
> +{
> +       unsigned int num = vq->packed.vring.num;
> +       u16 last_used, last_used_idx;
> +       bool used_wrap_counter;
> +       void *ret;
> +
> +       START_USE(vq);
> +
> +       if (unlikely(vq->broken)) {
> +               END_USE(vq);
> +               return NULL;
> +       }
> +
> +       last_used_idx = vq->last_used_idx;
> +       used_wrap_counter = packed_used_wrap_counter(last_used_idx);
> +       last_used = packed_last_used(last_used_idx);
> +
> +       if (vq->batch_last.id == UINT_MAX) {
> +               if (!more_used_packed_in_order(vq)) {
> +                       pr_debug("No more buffers in queue\n");
> +                       END_USE(vq);
> +                       return NULL;
> +               }
> +               /* Only get used elements after they have been exposed by host. */
> +               virtio_rmb(vq->weak_barriers);
> +               vq->batch_last.id =
> +                       le16_to_cpu(vq->packed.vring.desc[last_used].id);
> +               vq->batch_last.len =
> +                       le32_to_cpu(vq->packed.vring.desc[last_used].len);
> +       }
> +
> +       if (vq->batch_last.id == last_used) {
> +               vq->batch_last.id = UINT_MAX;
> +               *len = vq->batch_last.len;
> +       } else {
> +               *len = vq->packed.desc_state[last_used].total_len;
> +       }
> +
> +       if (unlikely(last_used >= num)) {
> +               BAD_RING(vq, "id %u out of range\n", last_used);
> +               return NULL;
> +       }
> +       if (unlikely(!vq->packed.desc_state[last_used].data)) {
> +               BAD_RING(vq, "id %u is not a head!\n", last_used);
> +               return NULL;
> +       }
> +
> +       /* detach_buf_packed clears data, so grab it now. */
> +       ret = vq->packed.desc_state[last_used].data;
> +       detach_buf_packed_in_order(vq, last_used, ctx);
> +
> +       update_last_used_idx_packed(vq, last_used, last_used,
> +                                   used_wrap_counter);
> +
> +       LAST_ADD_TIME_INVALID(vq);
> +
> +       END_USE(vq);
> +       return ret;
> +}
> +
>  static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
>                                           unsigned int *len,
>                                           void **ctx)
>  {
> +       unsigned int num = vq->packed.vring.num;
>         u16 last_used, id, last_used_idx;
>         bool used_wrap_counter;
>         void *ret;
> @@ -1826,7 +2181,7 @@ static void *virtqueue_get_buf_ctx_packed(struct vring_virtqueue *vq,
>         id = le16_to_cpu(vq->packed.vring.desc[last_used].id);
>         *len = le32_to_cpu(vq->packed.vring.desc[last_used].len);
>
> -       if (unlikely(id >= vq->packed.vring.num)) {
> +       if (unlikely(id >= num)) {
>                 BAD_RING(vq, "id %u out of range\n", id);
>                 return NULL;
>         }
> @@ -1967,7 +2322,10 @@ static void *virtqueue_detach_unused_buf_packed(struct vring_virtqueue *vq)
>                         continue;
>                 /* detach_buf clears data, so grab it now. */
>                 buf = vq->packed.desc_state[i].data;
> -               detach_buf_packed(vq, i, NULL);
> +               if (virtqueue_is_in_order(vq))
> +                       detach_buf_packed_in_order(vq, i, NULL);
> +               else
> +                       detach_buf_packed(vq, i, NULL);
>                 END_USE(vq);
>                 return buf;
>         }
> @@ -1993,6 +2351,8 @@ static struct vring_desc_extra *vring_alloc_desc_extra(unsigned int num)
>         for (i = 0; i < num - 1; i++)
>                 desc_extra[i].next = i + 1;
>
> +       desc_extra[num - 1].next = 0;
> +
>         return desc_extra;
>  }
>
> @@ -2124,10 +2484,17 @@ static void virtqueue_vring_attach_packed(struct vring_virtqueue *vq,
>  {
>         vq->packed = *vring_packed;
>
> -       /* Put everything in free lists. */
> -       vq->free_head = 0;
> +       if (virtqueue_is_in_order(vq)) {
> +               vq->batch_last.id = UINT_MAX;
> +       } else {
> +               /*
> +                * Put everything in free lists. Note that
> +                * next_avail_idx is sufficient with IN_ORDER so
> +                * free_head is unused.
> +                */
> +               vq->free_head = 0;
> +       }
>  }
> -
>  static void virtqueue_reset_packed(struct vring_virtqueue *vq)
>  {
>         memset(vq->packed.vring.device, 0, vq->packed.event_size_in_bytes);
> @@ -2172,13 +2539,14 @@ static struct virtqueue *__vring_new_virtqueue_packed(unsigned int index,
>  #else
>         vq->broken = false;
>  #endif
> -       vq->layout = PACKED;
>         vq->map = map;
>         vq->use_map_api = vring_use_map_api(vdev);
>
>         vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
>                 !context;
>         vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
> +       vq->layout = virtio_has_feature(vdev, VIRTIO_F_IN_ORDER) ?
> +                    PACKED_IN_ORDER : PACKED;
>
>         if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
>                 vq->weak_barriers = false;
> @@ -2288,9 +2656,39 @@ static const struct virtqueue_ops packed_ops = {
>         .reset = virtqueue_reset_packed,
>  };
>
> +static const struct virtqueue_ops split_in_order_ops = {
> +       .add = virtqueue_add_split,
> +       .get = virtqueue_get_buf_ctx_split_in_order,
> +       .kick_prepare = virtqueue_kick_prepare_split,
> +       .disable_cb = virtqueue_disable_cb_split,
> +       .enable_cb_delayed = virtqueue_enable_cb_delayed_split,
> +       .enable_cb_prepare = virtqueue_enable_cb_prepare_split,
> +       .poll = virtqueue_poll_split,
> +       .detach_unused_buf = virtqueue_detach_unused_buf_split,
> +       .more_used = more_used_split_in_order,
> +       .resize = virtqueue_resize_split,
> +       .reset = virtqueue_reset_split,
> +};
> +
> +static const struct virtqueue_ops packed_in_order_ops = {
> +       .add = virtqueue_add_packed_in_order,
> +       .get = virtqueue_get_buf_ctx_packed_in_order,
> +       .kick_prepare = virtqueue_kick_prepare_packed,
> +       .disable_cb = virtqueue_disable_cb_packed,
> +       .enable_cb_delayed = virtqueue_enable_cb_delayed_packed,
> +       .enable_cb_prepare = virtqueue_enable_cb_prepare_packed,
> +       .poll = virtqueue_poll_packed,
> +       .detach_unused_buf = virtqueue_detach_unused_buf_packed,
> +       .more_used = more_used_packed_in_order,
> +       .resize = virtqueue_resize_packed,
> +       .reset = virtqueue_reset_packed,
> +};
> +
>  static const struct virtqueue_ops *const all_ops[VQ_TYPE_MAX] = {
>         [SPLIT] = &split_ops,
> -       [PACKED] = &packed_ops
> +       [PACKED] = &packed_ops,
> +       [SPLIT_IN_ORDER] = &split_in_order_ops,
> +       [PACKED_IN_ORDER] = &packed_in_order_ops,
>  };
>
>  static int virtqueue_disable_and_recycle(struct virtqueue *_vq,
> @@ -2346,6 +2744,12 @@ static int virtqueue_enable_after_reset(struct virtqueue *_vq)
>         case PACKED:                                                    \
>                 ret = all_ops[PACKED]->op(vq, ##__VA_ARGS__);           \
>                 break;                                                  \
> +       case SPLIT_IN_ORDER:                                            \
> +               ret = all_ops[SPLIT_IN_ORDER]->op(vq, ##__VA_ARGS__);   \
> +               break;                                                  \
> +       case PACKED_IN_ORDER:                                           \
> +               ret = all_ops[PACKED_IN_ORDER]->op(vq, ##__VA_ARGS__);  \
> +               break;                                                  \
>         default:                                                        \
>                 BUG();                                                  \
>                 break;                                                  \
> @@ -2362,10 +2766,16 @@ static int virtqueue_enable_after_reset(struct virtqueue *_vq)
>         case PACKED:                                    \
>                 all_ops[PACKED]->op(vq, ##__VA_ARGS__); \
>                 break;                                  \
> -       default:                                        \
> -               BUG();                                  \
> -               break;                                  \
> -       }                                               \
> +       case SPLIT_IN_ORDER:                                            \
> +               all_ops[SPLIT_IN_ORDER]->op(vq, ##__VA_ARGS__); \
> +               break;                                                  \
> +       case PACKED_IN_ORDER:                                           \
> +               all_ops[PACKED_IN_ORDER]->op(vq, ##__VA_ARGS__);        \
> +               break;                                                  \
> +       default:                                                        \
> +               BUG();                                                  \
> +               break;                                                  \
> +       }                                                               \
>  })
>
>  static inline int virtqueue_add(struct virtqueue *_vq,
> @@ -3081,6 +3491,8 @@ void vring_transport_features(struct virtio_device *vdev)
>                         break;
>                 case VIRTIO_F_NOTIFICATION_DATA:
>                         break;
> +               case VIRTIO_F_IN_ORDER:
> +                       break;
>                 default:
>                         /* We don't understand this bit. */
>                         __virtio_clear_bit(vdev, i);
> --
> 2.31.1
>

Re: [PATCH V9 00/19] virtio_ring in order support

[Jason Wang]

On Tue, Nov 25, 2025 at 3:08 PM Jason Wang <jasowang@redhat.com> wrote:
>
> Hello all:
>
> This sereis tries to implement the VIRTIO_F_IN_ORDER to
> virtio_ring. This is done by introducing virtqueue ops so we can
> implement separate helpers for different virtqueue layout/features
> then the in-order were implemented on top.
>
> Tests shows 2%-19% imporvment with packed virtqueue PPS with KVM guest
> vhost-net/testpmd on the host.
>
> Changes since V8:
>
> - Rebase to vhost.git linux-next branch
> - Fix issue when indirect descriptor is disabled
> - use UINT_MAX instead of vq->num to indicate no batching is pending
> - Tweak the comments for used_entry
> - Various coding style tweaks and typo fixes
> - Try to use in_order varianet in in_order ops

Michael, any chance to make this for this merge window?

Thanks

Re: [PATCH V9 00/19] virtio_ring in order support

[Michael S. Tsirkin]

On Tue, Dec 23, 2025 at 09:18:47AM +0800, Jason Wang wrote:
> On Tue, Nov 25, 2025 at 3:08 PM Jason Wang <jasowang@redhat.com> wrote:
> >
> > Hello all:
> >
> > This sereis tries to implement the VIRTIO_F_IN_ORDER to
> > virtio_ring. This is done by introducing virtqueue ops so we can
> > implement separate helpers for different virtqueue layout/features
> > then the in-order were implemented on top.
> >
> > Tests shows 2%-19% imporvment with packed virtqueue PPS with KVM guest
> > vhost-net/testpmd on the host.
> >
> > Changes since V8:
> >
> > - Rebase to vhost.git linux-next branch
> > - Fix issue when indirect descriptor is disabled
> > - use UINT_MAX instead of vq->num to indicate no batching is pending
> > - Tweak the comments for used_entry
> > - Various coding style tweaks and typo fixes
> > - Try to use in_order varianet in in_order ops
> 
> Michael, any chance to make this for this merge window?
> 
> Thanks

want to fix last minor comments?

Re: [PATCH V9 00/19] virtio_ring in order support

[Jason Wang]

On Tue, Dec 23, 2025 at 9:29 AM Michael S. Tsirkin <mst@redhat.com> wrote:
>
> On Tue, Dec 23, 2025 at 09:18:47AM +0800, Jason Wang wrote:
> > On Tue, Nov 25, 2025 at 3:08 PM Jason Wang <jasowang@redhat.com> wrote:
> > >
> > > Hello all:
> > >
> > > This sereis tries to implement the VIRTIO_F_IN_ORDER to
> > > virtio_ring. This is done by introducing virtqueue ops so we can
> > > implement separate helpers for different virtqueue layout/features
> > > then the in-order were implemented on top.
> > >
> > > Tests shows 2%-19% imporvment with packed virtqueue PPS with KVM guest
> > > vhost-net/testpmd on the host.
> > >
> > > Changes since V8:
> > >
> > > - Rebase to vhost.git linux-next branch
> > > - Fix issue when indirect descriptor is disabled
> > > - use UINT_MAX instead of vq->num to indicate no batching is pending
> > > - Tweak the comments for used_entry
> > > - Various coding style tweaks and typo fixes
> > > - Try to use in_order varianet in in_order ops
> >
> > Michael, any chance to make this for this merge window?
> >
> > Thanks
>
> want to fix last minor comments?

Yes, I will post a new version.

Thanks

>