Re: [PATCH v7 09/26] virtio_ring: split: implement virtqueue_reset_vring_split()

On Thu, 10 Mar 2022 07:17:09 -0500, "Michael S. Tsirkin" [off-list ref] wrote:

On Thu, Mar 10, 2022 at 04:14:16PM +0800, Xuan Zhuo wrote:

quoted

On Thu, 10 Mar 2022 03:07:22 -0500, "Michael S. Tsirkin" [off-list ref] wrote:

quoted

On Thu, Mar 10, 2022 at 03:17:03PM +0800, Xuan Zhuo wrote:

quoted

On Thu, 10 Mar 2022 02:00:39 -0500, "Michael S. Tsirkin" [off-list ref] wrote:

quoted

On Tue, Mar 08, 2022 at 08:35:01PM +0800, Xuan Zhuo wrote:

quoted

virtio ring supports reset.

Queue reset is divided into several stages.

1. notify device queue reset
2. vring release
3. attach new vring
4. notify device queue re-enable

After the first step is completed, the vring reset operation can be
performed. If the newly set vring num does not change, then just reset
the vq related value.

Otherwise, the vring will be released and the vring will be reallocated.
And the vring will be attached to the vq. If this process fails, the
function will exit, and the state of the vq will be the vring release
state. You can call this function again to reallocate the vring.

In addition, vring_align, may_reduce_num are necessary for reallocating
vring, so they are retained when creating vq.

Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
---
 drivers/virtio/virtio_ring.c | 69 ++++++++++++++++++++++++++++++++++++
 1 file changed, 69 insertions(+)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index e0422c04c903..148fb1fd3d5a 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c

@@ -158,6 +158,12 @@ struct vring_virtqueue {
 			/* DMA address and size information */
 			dma_addr_t queue_dma_addr;
 			size_t queue_size_in_bytes;
+
+			/* The parameters for creating vrings are reserved for
+			 * creating new vrings when enabling reset queue.
+			 */
+			u32 vring_align;
+			bool may_reduce_num;
 		} split;

 		/* Available for packed ring */

@@ -217,6 +223,12 @@ struct vring_virtqueue {
 #endif
 };

+static void vring_free(struct virtqueue *vq);
+static void __vring_virtqueue_init_split(struct vring_virtqueue *vq,
+					 struct virtio_device *vdev);
+static int __vring_virtqueue_attach_split(struct vring_virtqueue *vq,
+					  struct virtio_device *vdev,
+					  struct vring vring);

 /*
  * Helpers.

@@ -1012,6 +1024,8 @@ static struct virtqueue *vring_create_virtqueue_split(
 		return NULL;
 	}

+	to_vvq(vq)->split.vring_align = vring_align;
+	to_vvq(vq)->split.may_reduce_num = may_reduce_num;
 	to_vvq(vq)->split.queue_dma_addr = vring.dma_addr;
 	to_vvq(vq)->split.queue_size_in_bytes = vring.queue_size_in_bytes;
 	to_vvq(vq)->we_own_ring = true;

@@ -1019,6 +1033,59 @@ static struct virtqueue *vring_create_virtqueue_split(
 	return vq;
 }

+static int virtqueue_reset_vring_split(struct virtqueue *_vq, u32 num)
+{
+	struct vring_virtqueue *vq = to_vvq(_vq);
+	struct virtio_device *vdev = _vq->vdev;
+	struct vring_split vring;
+	int err;
+
+	if (num > _vq->num_max)
+		return -E2BIG;
+
+	switch (vq->vq.reset) {
+	case VIRTIO_VQ_RESET_STEP_NONE:
+		return -ENOENT;
+
+	case VIRTIO_VQ_RESET_STEP_VRING_ATTACH:
+	case VIRTIO_VQ_RESET_STEP_DEVICE:
+		if (vq->split.vring.num == num || !num)
+			break;
+
+		vring_free(_vq);
+
+		fallthrough;
+
+	case VIRTIO_VQ_RESET_STEP_VRING_RELEASE:
+		if (!num)
+			num = vq->split.vring.num;
+
+		err = vring_create_vring_split(&vring, vdev,
+					       vq->split.vring_align,
+					       vq->weak_barriers,
+					       vq->split.may_reduce_num, num);
+		if (err)
+			return -ENOMEM;
+
+		err = __vring_virtqueue_attach_split(vq, vdev, vring.vring);
+		if (err) {
+			vring_free_queue(vdev, vring.queue_size_in_bytes,
+					 vring.queue,
+					 vring.dma_addr);
+			return -ENOMEM;
+		}
+
+		vq->split.queue_dma_addr = vring.dma_addr;
+		vq->split.queue_size_in_bytes = vring.queue_size_in_bytes;
+	}
+
+	__vring_virtqueue_init_split(vq, vdev);
+	vq->we_own_ring = true;
+	vq->vq.reset = VIRTIO_VQ_RESET_STEP_VRING_ATTACH;
+
+	return 0;
+}
+

I kind of dislike this state machine.

Hacks like special-casing num = 0 to mean "reset" are especially
confusing.

I'm removing it. I'll say in the function description that this function is
currently only called when vq has been reset. I'm no longer checking it based on
state.

quoted

And as Jason points out, when we want a resize then yes this currently
implies reset but that is an implementation detail.

There should be a way to just make these cases separate functions
and then use them to compose consistent external APIs.

Yes, virtqueue_resize_split() is fine for ethtool -G.

But in the case of AF_XDP, just execute reset to free the buffer. The name
virtqueue_reset_vring_split() I think can cover both cases. Or we use two apis
to handle both scenarios?

Or can anyone think of a better name. ^_^

Thanks.

I'd say resize should be called resize and reset should be called reset.

OK, I'll change it to resize here.

But I want to know that when I implement virtio-net to support AF_XDP, its
requirement is to release all submitted buffers. Then should I add a new api
such as virtqueue_reset_vring()?

Sounds like a reasonable name.

quoted

quoted

The big issue is a sane API for resize. Ideally it would resubmit
buffers which did not get used. Question is what to do
about buffers which don't fit (if ring has been downsized)?
Maybe a callback that will handle them?
And then what? Queue them up and readd later? Drop?
If we drop we should drop from the head not the tail ...

It's a good idea, let's implement it later.

Thanks.

Well ... not sure how you are going to support resize
if you don't know what to do with buffers that were
in the ring.

The current solution is to call virtqueue_detach_unused_buf() to release buffers
before resize ring.

Thanks.