[v3] vhost: batch used descs chains write-back with packed ring

  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.coquelin@redhat.com>
---
 lib/librte_vhost/virtio_net.c | 19 +++++++++++++++++--
 1 file changed, 17 insertions(+), 2 deletions(-)
  

On Thu, Dec 20, 2018 at 11:00:22AM +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.coquelin@redhat.com>
> ---
>  lib/librte_vhost/virtio_net.c | 19 +++++++++++++++++--
>  1 file changed, 17 insertions(+), 2 deletions(-)

Reviewed-by: Tiwei Bie <tiwei.bie@intel.com>
  

On Thu, Dec 20, 2018 at 11:00:22AM +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.coquelin@redhat.com>
> ---
>  lib/librte_vhost/virtio_net.c | 19 +++++++++++++++++--
>  1 file changed, 17 insertions(+), 2 deletions(-)
> 
> diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
> index 8c657a101..66ccd3c35 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 +147,15 @@ flush_shadow_used_ring_packed(struct virtio_net *dev,
>  		}
>  	}
>  
> +	vq->desc_packed[head_idx].flags = head_flags;
> +
>  	rte_smp_wmb();
> +
> +	vhost_log_cache_used_vring(dev, vq,
> +				head_idx *
> +				sizeof(struct vring_packed_desc),
> +				sizeof(struct vring_packed_desc));
> +
>  	vq->shadow_used_idx = 0;
>  	vhost_log_cache_sync(dev, vq);

How about moving rte_smp_wmb into logging functions?
This way it's free with log disabled even on arm...

>  }
> -- 
> 2.17.2
  

On 12/20/18 3:30 PM, Michael S. Tsirkin wrote:
> On Thu, Dec 20, 2018 at 11:00:22AM +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.coquelin@redhat.com>
>> ---
>>   lib/librte_vhost/virtio_net.c | 19 +++++++++++++++++--
>>   1 file changed, 17 insertions(+), 2 deletions(-)
>>
>> diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
>> index 8c657a101..66ccd3c35 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 +147,15 @@ flush_shadow_used_ring_packed(struct virtio_net *dev,
>>   		}
>>   	}
>>   
>> +	vq->desc_packed[head_idx].flags = head_flags;
>> +
>>   	rte_smp_wmb();
>> +
>> +	vhost_log_cache_used_vring(dev, vq,
>> +				head_idx *
>> +				sizeof(struct vring_packed_desc),
>> +				sizeof(struct vring_packed_desc));
>> +
>>   	vq->shadow_used_idx = 0;
>>   	vhost_log_cache_sync(dev, vq);
> 
> How about moving rte_smp_wmb into logging functions?
> This way it's free with log disabled even on arm...

That's what I initially suggested in my reply to v2.
Problem is that in split ring case, we already have a barrier before
cache sync, and we need it even if logging is disabled.

But I think you are right, it might be better to have the barrier twice
in split ring case when logging is enabled and none for packed ring when
logging is disabled.

I'll post a v4.

Thanks,
Maxime

>>   }
>> -- 
>> 2.17.2