On 12/20/18 9:49 AM, Maxime Coquelin wrote:
On 12/20/18 5:44 AM, Tiwei Bie wrote:
On Wed, Dec 19, 2018 at 10:29:52AM +0100, Maxime Coquelin wrote:
Instead of writing back descriptors chains in order, let's
write the first chain flags last in order to improve batching.
With Kernel's pktgen benchmark, ~3% performance gain is measured.
Signed-off-by: Maxime Coquelin <maxime.coque...@redhat.com>
---
V2:
Revert back to initial implementation to have a write
barrier before every descs flags store, but still
store first desc flags last. (Missing barrier reported
by Ilya)
lib/librte_vhost/virtio_net.c | 19 ++++++++++++++++---
1 file changed, 16 insertions(+), 3 deletions(-)
diff --git a/lib/librte_vhost/virtio_net.c
b/lib/librte_vhost/virtio_net.c
index 8c657a101..de436af79 100644
--- a/lib/librte_vhost/virtio_net.c
+++ b/lib/librte_vhost/virtio_net.c
@@ -97,6 +97,8 @@ flush_shadow_used_ring_packed(struct virtio_net *dev,
{
int i;
uint16_t used_idx = vq->last_used_idx;
+ uint16_t head_idx = vq->last_used_idx;
+ uint16_t head_flags = 0;
/* Split loop in two to save memory barriers */
for (i = 0; i < vq->shadow_used_idx; i++) {
@@ -126,12 +128,17 @@ flush_shadow_used_ring_packed(struct virtio_net
*dev,
flags &= ~VRING_DESC_F_AVAIL;
}
- vq->desc_packed[vq->last_used_idx].flags = flags;
+ if (i > 0) {
+ vq->desc_packed[vq->last_used_idx].flags = flags;
- vhost_log_cache_used_vring(dev, vq,
+ vhost_log_cache_used_vring(dev, vq,
vq->last_used_idx *
sizeof(struct vring_packed_desc),
sizeof(struct vring_packed_desc));
+ } else {
+ head_idx = vq->last_used_idx;
+ head_flags = flags;
+ }
vq->last_used_idx += vq->shadow_used_packed[i].count;
if (vq->last_used_idx >= vq->size) {
@@ -140,7 +147,13 @@ flush_shadow_used_ring_packed(struct virtio_net
*dev,
}
}
- rte_smp_wmb();
+ vq->desc_packed[head_idx].flags = head_flags;
+
+ vhost_log_cache_used_vring(dev, vq,
+ vq->last_used_idx *
Should be head_idx.
Oh yes, thanks for spotting this.
+ sizeof(struct vring_packed_desc),
+ sizeof(struct vring_packed_desc));
+
vq->shadow_used_idx = 0;
A wmb() is needed before log_cache_sync?
I think you're right, I was wrong but thought we had a barrier in cache
sync function.
That's not very important for x86, but I think it should be preferable
to do it in vhost_log_cache_sync(), if logging is enabled.
What do you think?
I'll keep it in this function for now, as I think we cannot remove the
one in the split variant so it would mean having two barriers in that
case.
vhost_log_cache_sync(dev, vq);
}
--
2.17.2
