[dpdk-dev] [PATCH v2 3/3] net/virtio: improve batching in mergeable path

[Maxime Coquelin]

On 12/19/18 10:47 AM, Jens Freimann wrote:
> On Tue, Dec 11, 2018 at 02:48:04PM +0100, Maxime Coquelin wrote:
>> This patch improves both descriptors dequeue and refill,
>> by using the same bathing strategy as done in in-order path.
> 
> s/bathing/batching/
> 
>>
>> Signed-off-by: Maxime Coquelin <maxime.coquelin at redhat.com>
>> ---
>> drivers/net/virtio/virtio_rxtx.c | 237 ++++++++++++++++---------------
>> 1 file changed, 126 insertions(+), 111 deletions(-)
>>
>> diff --git a/drivers/net/virtio/virtio_rxtx.c 
>> b/drivers/net/virtio/virtio_rxtx.c
>> index ebe5c74b5..59bcac2f7 100644
>> --- a/drivers/net/virtio/virtio_rxtx.c
>> +++ b/drivers/net/virtio/virtio_rxtx.c
>> @@ -267,41 +267,42 @@ virtqueue_enqueue_refill_inorder(struct 
>> virtqueue *vq,
>> }
>>
>> static inline int
>> -virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf 
>> *cookie)
>> +virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf 
>> **cookie,
>> +                uint16_t num)
>> {
>>     struct vq_desc_extra *dxp;
>>     struct virtio_hw *hw = vq->hw;
>> -    struct vring_desc *start_dp;
>> -    uint16_t needed = 1;
>> -    uint16_t head_idx, idx;
>> +    struct vring_desc *start_dp = vq->vq_ring.desc;
>> +    uint16_t idx, i;
>>
>>     if (unlikely(vq->vq_free_cnt == 0))
>>         return -ENOSPC;
>> -    if (unlikely(vq->vq_free_cnt < needed))
>> +    if (unlikely(vq->vq_free_cnt < num))
>>         return -EMSGSIZE;
>>
>> -    head_idx = vq->vq_desc_head_idx;
>> -    if (unlikely(head_idx >= vq->vq_nentries))
>> +    if (unlikely(vq->vq_desc_head_idx >= vq->vq_nentries))
>>         return -EFAULT;
>>
>> -    idx = head_idx;
>> -    dxp = &vq->vq_descx[idx];
>> -    dxp->cookie = (void *)cookie;
>> -    dxp->ndescs = needed;
>> +    for (i = 0; i < num; i++) {
>> +        idx = vq->vq_desc_head_idx;
>> +        dxp = &vq->vq_descx[idx];
>> +        dxp->cookie = (void *)cookie[i];
> 
> I think code is safe as it is, but should we check if cookie actually 
> points to something?

I don't think it is needed, we check in virtqueue_dequeue_burst_rx that
cookie isn't NULL before using it.

Thanks,
Maxime

> I tested this patch and saw the same performance improvement, so
> 
> Tested-by: Jens Freimann <jfreimann at redhat.com>
> Reviewed-by: Jens Freimann <jfreimann at redhat.com>
> 
> regards,
> Jens