[dpdk-dev] [PATCH v2 03/16] vhost: add burst enqueue function for packed ring

[Liu, Yong]

Thanks, Gavin. My comments are inline.

> -----Original Message-----
> From: Gavin Hu (Arm Technology China) [mailto:Gavin.Hu at arm.com]
> Sent: Monday, September 23, 2019 7:09 PM
> To: Liu, Yong <yong.liu at intel.com>; maxime.coquelin at redhat.com; Bie, Tiwei
> <tiwei.bie at intel.com>; Wang, Zhihong <zhihong.wang at intel.com>
> Cc: dev at dpdk.org; nd <nd at arm.com>
> Subject: RE: [dpdk-dev] [PATCH v2 03/16] vhost: add burst enqueue function
> for packed ring
> 
> Hi Marvin,
> 
> Is it possible to vectorize the processing?
> Other comments inline:
> /Gavin

Gavin,
According to our experiment, only vectorize some parts in [ed]nqueue function can't benefit performance.
Functions like vhost_iova_to_vva and virtio_enqueue_offload can't be easily vectorized as they are full of judgment conditions.

Thanks,
Marvin

> > -----Original Message-----
> > From: dev <dev-bounces at dpdk.org> On Behalf Of Marvin Liu
> > Sent: Friday, September 20, 2019 12:37 AM
> > To: maxime.coquelin at redhat.com; tiwei.bie at intel.com;
> > zhihong.wang at intel.com
> > Cc: dev at dpdk.org; Marvin Liu <yong.liu at intel.com>
> > Subject: [dpdk-dev] [PATCH v2 03/16] vhost: add burst enqueue function
> for
> > packed ring
> >
> > Burst enqueue function will first check whether descriptors are cache
> > aligned. It will also check prerequisites in the beginning. Burst
> > enqueue function not support chained mbufs, single packet enqueue
> > function will handle it.
> >
> > 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 5074226f0..67889c80a 100644
> > --- a/lib/librte_vhost/vhost.h
> > +++ b/lib/librte_vhost/vhost.h
> > @@ -39,6 +39,9 @@
> >
> >  #define VHOST_LOG_CACHE_NR 32
> >
> > +#define PACKED_DESCS_BURST (RTE_CACHE_LINE_SIZE / \
> > +			    sizeof(struct vring_packed_desc))
> > +
> >  #ifdef SUPPORT_GCC_UNROLL_PRAGMA
> >  #define PRAGMA_PARAM "GCC unroll 4"
> >  #endif
> > @@ -57,6 +60,8 @@
> >  #define UNROLL_PRAGMA(param) do {} while(0);
> >  #endif
> >
> > +#define PACKED_BURST_MASK (PACKED_DESCS_BURST - 1)
> > +
> >  /**
> >   * 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 2b5c47145..c664b27c5 100644
> > --- a/lib/librte_vhost/virtio_net.c
> > +++ b/lib/librte_vhost/virtio_net.c
> > @@ -895,6 +895,84 @@ virtio_dev_rx_split(struct virtio_net *dev, struct
> > vhost_virtqueue *vq,
> >  	return pkt_idx;
> >  }
> >
> > +static __rte_unused __rte_always_inline int
> I remember "__rte_always_inline" should start at the first and separate
> line, otherwise you will get a style issue.
> /Gavin
> > +virtio_dev_rx_burst_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_addrs[PACKED_DESCS_BURST];
> > +	struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_DESCS_BURST];
> > +	uint32_t buf_offset = dev->vhost_hlen;
> > +	uint64_t lens[PACKED_DESCS_BURST];
> > +
> > +	uint16_t i;
> > +
> > +	if (unlikely(avail_idx & PACKED_BURST_MASK))
> > +		return -1;
> > +
> > +	UNROLL_PRAGMA(PRAGMA_PARAM)
> > +	for (i = 0; i < PACKED_DESCS_BURST; i++) {
> > +		if (unlikely(pkts[i]->next != NULL))
> > +			return -1;
> > +		if (unlikely(!desc_is_avail(&descs[avail_idx + i],
> > +					    wrap_counter)))
> > +			return -1;
> > +	}
> > +
> > +	rte_smp_rmb();
> > +
> > +	UNROLL_PRAGMA(PRAGMA_PARAM)
> > +	for (i = 0; i < PACKED_DESCS_BURST; i++)
> > +		lens[i] = descs[avail_idx + i].len;
> Looks like the code is a strong candidate for vectorization.
> > +
> > +	UNROLL_PRAGMA(PRAGMA_PARAM)
> > +	for (i = 0; i < PACKED_DESCS_BURST; i++) {
> > +		if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
> > +			return -1;
> > +	}
> > +
> > +	UNROLL_PRAGMA(PRAGMA_PARAM)
> > +	for (i = 0; i < PACKED_DESCS_BURST; i++)
> > +		desc_addrs[i] = vhost_iova_to_vva(dev, vq,
> > +						 descs[avail_idx + i].addr,
> > +						 &lens[i],
> > +						 VHOST_ACCESS_RW);
> > +	UNROLL_PRAGMA(PRAGMA_PARAM)
> > +	for (i = 0; i < PACKED_DESCS_BURST; i++) {
> > +		if (unlikely(lens[i] != descs[avail_idx + i].len))
> > +			return -1;
> > +	}
> > +
> > +	UNROLL_PRAGMA(PRAGMA_PARAM)
> > +	for (i = 0; i < PACKED_DESCS_BURST; i++) {
> > +		rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
> > +		hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)desc_addrs[i];
> > +		lens[i] = pkts[i]->pkt_len + dev->vhost_hlen;
> > +	}
> > +
> > +	UNROLL_PRAGMA(PRAGMA_PARAM)
> > +	for (i = 0; i < PACKED_DESCS_BURST; i++)
> > +		virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
> > +
> A store barrier here is missing, last_avail_idx may be observed before the
> above enqueue completion on weak memory order architectures.
> For x86, a compiler barrier is also required.
> 
Thanks a lot for point out. I guess your mention is that need to add barrier between memcpy and enqueue.
last_avail_idx is just local variable, no barrier is need to protect it.

> > +	vq->last_avail_idx += PACKED_DESCS_BURST;
> > +	if (vq->last_avail_idx >= vq->size) {
> > +		vq->last_avail_idx -= vq->size;
> > +		vq->avail_wrap_counter ^= 1;
> > +	}
> > +
> > +	UNROLL_PRAGMA(PRAGMA_PARAM)
> > +	for (i = 0; i < PACKED_DESCS_BURST; i++) {
> > +		rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
> > +			   rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
> > +			   pkts[i]->pkt_len);
> > +	}
> > +
> > +	return 0;
> > +}
> > +
> >  static __rte_unused int16_t
> >  virtio_dev_rx_single_packed(struct virtio_net *dev, struct
> vhost_virtqueue
> > *vq,
> >  	struct rte_mbuf *pkt)
> > --
> > 2.17.1