[dpdk-dev] [RFC PATCH 02/13] add vhost packed ring fast enqueue function

[Liu, Yong]

Thank, Ilya.

> -----Original Message-----
> From: Ilya Maximets [mailto:i.maximets at samsung.com]
> Sent: Monday, July 08, 2019 7:38 PM
> To: Liu, Yong <yong.liu at intel.com>; Bie, Tiwei <tiwei.bie at intel.com>;
> maxime.coquelin at redhat.com; dev at dpdk.org
> Subject: Re: [dpdk-dev] [RFC PATCH 02/13] add vhost packed ring fast enqueue
> function
> 
> On 08.07.2019 20:13, Marvin Liu wrote:
> > In fast enqueue function, will first check whether descriptors are
> > cache aligned. Fast enqueue function will check prerequisites in the
> > beginning. Fast enqueue function do not support chained mbufs, normal
> > function will handle that.
> >
> > Signed-off-by: Marvin Liu <yong.liu at intel.com>
> >
> > diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
> > index 884befa85..f24026acd 100644
> > --- a/lib/librte_vhost/vhost.h
> > +++ b/lib/librte_vhost/vhost.h
> > @@ -39,6 +39,8 @@
> >
> >  #define VHOST_LOG_CACHE_NR 32
> >
> > +/* Used in fast packed ring functions */
> > +#define PACKED_DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct
> vring_packed_desc))
> >  /**
> >   * Structure contains buffer address, length and descriptor index
> >   * from vring to do scatter RX.
> > diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
> > index 003aec1d4..b877510da 100644
> > --- a/lib/librte_vhost/virtio_net.c
> > +++ b/lib/librte_vhost/virtio_net.c
> > @@ -897,6 +897,115 @@ virtio_dev_rx_split(struct virtio_net *dev, struct
> vhost_virtqueue *vq,
> >  	return pkt_idx;
> >  }
> >
> > +static __rte_always_inline uint16_t
> > +virtio_dev_rx_fast_packed(struct virtio_net *dev, struct vhost_virtqueue
> *vq,
> > +		struct rte_mbuf **pkts)
> > +{
> > +	bool wrap_counter = vq->avail_wrap_counter;
> > +	struct vring_packed_desc *descs = vq->desc_packed;
> > +	uint16_t avail_idx = vq->last_avail_idx;
> > +	uint64_t desc_addr, desc_addr1, desc_addr2, desc_addr3, len, len1,
> > +		len2, len3;
> > +	struct virtio_net_hdr_mrg_rxbuf *hdr, *hdr1, *hdr2, *hdr3;
> > +	uint32_t buf_offset = dev->vhost_hlen;
> > +
> > +	if (unlikely(avail_idx & 0x3))
> > +		return -1;
> > +
> > +	if (unlikely(avail_idx < (vq->size - PACKED_DESC_PER_CACHELINE)))
> 
> Doe it makes sense to check this?
> If this condition is not 'true', all the code below will access incorrect
> memory.
> 
Sorry, this condition should be likely:)

> > +		rte_prefetch0((void *)(uintptr_t)&descs[avail_idx +
> > +			PACKED_DESC_PER_CACHELINE]);
> > +	else
> > +		rte_prefetch0((void *)(uintptr_t)&descs[0]);
> > +
> > +	if (unlikely((pkts[0]->next != NULL) |
> > +		(pkts[1]->next != NULL) |
> > +		(pkts[2]->next != NULL) |
> > +		(pkts[3]->next != NULL)))
> > +		return -1;
> > +
> > +	if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)) |
> > +		unlikely(!desc_is_avail(&descs[avail_idx + 1], wrap_counter)) |
> > +		unlikely(!desc_is_avail(&descs[avail_idx + 2], wrap_counter)) |
> > +		unlikely(!desc_is_avail(&descs[avail_idx + 3], wrap_counter)))
> > +		return 1;
> > +
> > +	rte_smp_rmb();
> > +
> > +	len = descs[avail_idx].len;
> > +	len1 = descs[avail_idx + 1].len;
> > +	len2 = descs[avail_idx + 2].len;
> > +	len3 = descs[avail_idx + 3].len;
> > +
> > +	if (unlikely((pkts[0]->pkt_len > (len - buf_offset)) |
> > +		(pkts[1]->pkt_len > (len1 - buf_offset)) |
> > +		(pkts[2]->pkt_len > (len2 - buf_offset)) |
> > +		(pkts[3]->pkt_len > (len3 - buf_offset))))
> > +		return -1;
> > +
> > +	desc_addr = vhost_iova_to_vva(dev, vq,
> > +			descs[avail_idx].addr,
> > +			&len,
> > +			VHOST_ACCESS_RW);
> > +
> > +	desc_addr1 = vhost_iova_to_vva(dev, vq,
> > +			descs[avail_idx + 1].addr,
> > +			&len1,
> > +			VHOST_ACCESS_RW);
> > +
> > +	desc_addr2 = vhost_iova_to_vva(dev, vq,
> > +			descs[avail_idx + 2].addr,
> > +			&len2,
> > +			VHOST_ACCESS_RW);
> > +
> > +	desc_addr3 = vhost_iova_to_vva(dev, vq,
> > +			descs[avail_idx + 3].addr,
> > +			&len3,
> > +			VHOST_ACCESS_RW);
> > +
> > +	if (unlikely((len != descs[avail_idx].len) |
> > +		(len1 != descs[avail_idx + 1].len) |
> > +		(len2 != descs[avail_idx + 2].len) |
> > +		(len3 != descs[avail_idx + 3].len)))
> > +		return -1;
> > +
> > +	hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr;
> > +	hdr1 = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr1;
> > +	hdr2 = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr2;
> > +	hdr3 = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr3;
> > +
> > +	virtio_enqueue_offload(pkts[0], &hdr->hdr);
> > +	virtio_enqueue_offload(pkts[1], &hdr1->hdr);
> > +	virtio_enqueue_offload(pkts[2], &hdr2->hdr);
> > +	virtio_enqueue_offload(pkts[3], &hdr3->hdr);
> > +
> > +	len = pkts[0]->pkt_len + dev->vhost_hlen;
> > +	len1 = pkts[1]->pkt_len + dev->vhost_hlen;
> > +	len2 = pkts[2]->pkt_len + dev->vhost_hlen;
> > +	len3 = pkts[3]->pkt_len + dev->vhost_hlen;
> > +
> > +	vq->last_avail_idx += PACKED_DESC_PER_CACHELINE;
> 
> The whole function assumes that PACKED_DESC_PER_CACHELINE equals to 4,
> but if it's not, it will not work correctly.

Thanks for point out this, I just noticed that power8 has 128 bytes cache line.
Main idea here is to unroll the loop and check conditions first.
PACKED_DESC_PER_CACHELINE can be removed and renamed like FAST_BURST_NUM which hardcoded to four.

Best regards,
Marvin