Re: [RFC 1/5] vhost: reorder used descriptors in a batch

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 




在 2022/7/21 16:43, Guo Zhi 写道:
Device may not use descriptors in order, for example, NIC and SCSI may
not call __vhost_add_used_n with buffers in order.  It's the task of
__vhost_add_used_n to order them.


I'm not sure this is ture. Having ooo descriptors is probably by design to have better performance.

This might be obvious for device that may have elevator or QOS stuffs.

I suspect the right thing to do here is, for the device that can't perform better in the case of IN_ORDER, let's simply not offer IN_ORDER (zerocopy or scsi). And for the device we know it can perform better, non-zercopy ethernet device we can do that.


  This commit reorder the buffers using
vq->heads, only the batch is begin from the expected start point and is
continuous can the batch be exposed to driver.  And only writing out a
single used ring for a batch of descriptors, according to VIRTIO 1.1
spec.


So this sounds more like a "workaround" of the device that can't consume buffer in order, I suspect it can help in performance.

More below.



Signed-off-by: Guo Zhi <qtxuning1999@xxxxxxxxxxx>
---
  drivers/vhost/vhost.c | 44 +++++++++++++++++++++++++++++++++++++++++--
  drivers/vhost/vhost.h |  3 +++
  2 files changed, 45 insertions(+), 2 deletions(-)

diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index 40097826c..e2e77e29f 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c
@@ -317,6 +317,7 @@ static void vhost_vq_reset(struct vhost_dev *dev,
  	vq->used_flags = 0;
  	vq->log_used = false;
  	vq->log_addr = -1ull;
+	vq->next_used_head_idx = 0;
  	vq->private_data = NULL;
  	vq->acked_features = 0;
  	vq->acked_backend_features = 0;
@@ -398,6 +399,8 @@ static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
  					  GFP_KERNEL);
  		if (!vq->indirect || !vq->log || !vq->heads)
  			goto err_nomem;
+
+		memset(vq->heads, 0, sizeof(*vq->heads) * dev->iov_limit);
  	}
  	return 0;
@@ -2374,12 +2377,49 @@ static int __vhost_add_used_n(struct vhost_virtqueue *vq,
  			    unsigned count)
  {
  	vring_used_elem_t __user *used;
+	struct vring_desc desc;
  	u16 old, new;
  	int start;
+	int begin, end, i;
+	int copy_n = count;
+
+	if (vhost_has_feature(vq, VIRTIO_F_IN_ORDER)) {


How do you guarantee that ids of heads are contiguous?


+		/* calculate descriptor chain length for each used buffer */


I'm a little bit confused about this comment, we have heads[i].len for this?


+		for (i = 0; i < count; i++) {
+			begin = heads[i].id;
+			end = begin;
+			vq->heads[begin].len = 0;


Does this work for e.g RX virtqueue?


+			do {
+				vq->heads[begin].len += 1;
+				if (unlikely(vhost_get_desc(vq, &desc, end))) {


Let's try hard to avoid more userspace copy here, it's the source of performance regression.

Thanks


+					vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
+					       end, vq->desc + end);
+					return -EFAULT;
+				}
+			} while ((end = next_desc(vq, &desc)) != -1);
+		}
+
+		count = 0;
+		/* sort and batch continuous used ring entry */
+		while (vq->heads[vq->next_used_head_idx].len != 0) {
+			count++;
+			i = vq->next_used_head_idx;
+			vq->next_used_head_idx = (vq->next_used_head_idx +
+						  vq->heads[vq->next_used_head_idx].len)
+						  % vq->num;
+			vq->heads[i].len = 0;
+		}
+		/* only write out a single used ring entry with the id corresponding
+		 * to the head entry of the descriptor chain describing the last buffer
+		 * in the batch.
+		 */
+		heads[0].id = i;
+		copy_n = 1;
+	}
start = vq->last_used_idx & (vq->num - 1);
  	used = vq->used->ring + start;
-	if (vhost_put_used(vq, heads, start, count)) {
+	if (vhost_put_used(vq, heads, start, copy_n)) {
  		vq_err(vq, "Failed to write used");
  		return -EFAULT;
  	}
@@ -2410,7 +2450,7 @@ int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
start = vq->last_used_idx & (vq->num - 1);
  	n = vq->num - start;
-	if (n < count) {
+	if (n < count && !vhost_has_feature(vq, VIRTIO_F_IN_ORDER)) {
  		r = __vhost_add_used_n(vq, heads, n);
  		if (r < 0)
  			return r;
diff --git a/drivers/vhost/vhost.h b/drivers/vhost/vhost.h
index d9109107a..7b2c0fbb5 100644
--- a/drivers/vhost/vhost.h
+++ b/drivers/vhost/vhost.h
@@ -107,6 +107,9 @@ struct vhost_virtqueue {
  	bool log_used;
  	u64 log_addr;
+ /* Sort heads in order */
+	u16 next_used_head_idx;
+
  	struct iovec iov[UIO_MAXIOV];
  	struct iovec iotlb_iov[64];
  	struct iovec *indirect;




[Index of Archives]     [KVM ARM]     [KVM ia64]     [KVM ppc]     [Virtualization Tools]     [Spice Development]     [Libvirt]     [Libvirt Users]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite Questions]     [Linux Kernel]     [Linux SCSI]     [XFree86]

  Powered by Linux