diff mbox

[dpdk-dev,v3,1/5] vhost: rewrite enqueue

Message ID 1471585430-125925-2-git-send-email-zhihong.wang@intel.com (mailing list archive)
State Superseded, archived
Delegated to: Yuanhan Liu
Headers show

Commit Message

Zhihong Wang Aug. 19, 2016, 5:43 a.m. UTC
This patch implements the vhost logic from scratch into a single function
designed for high performance and better maintainability.

---
Changes in v3:

 1. Rewrite enqueue and delete the obsolete in the same patch.

Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
---
 lib/librte_vhost/vhost_rxtx.c | 537 +++++++++++++-----------------------------
 1 file changed, 160 insertions(+), 377 deletions(-)

Comments

Maxime Coquelin Aug. 22, 2016, 9:35 a.m. UTC | #1
On 08/19/2016 07:43 AM, Zhihong Wang wrote:
> This patch implements the vhost logic from scratch into a single function
> designed for high performance and better maintainability.
>
> ---
> Changes in v3:
>
>  1. Rewrite enqueue and delete the obsolete in the same patch.
>
> Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
> ---
>  lib/librte_vhost/vhost_rxtx.c | 537 +++++++++++++-----------------------------
>  1 file changed, 160 insertions(+), 377 deletions(-)
>
> diff --git a/lib/librte_vhost/vhost_rxtx.c b/lib/librte_vhost/vhost_rxtx.c
> index 08a73fd..b09a9c3 100644
> --- a/lib/librte_vhost/vhost_rxtx.c
> +++ b/lib/librte_vhost/vhost_rxtx.c
> @@ -91,7 +91,7 @@ is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t qp_nb)
>  	return (is_tx ^ (idx & 1)) == 0 && idx < qp_nb * VIRTIO_QNUM;
>  }
>
> -static void
> +static inline void __attribute__((always_inline))
>  virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
>  {
>  	if (m_buf->ol_flags & PKT_TX_L4_MASK) {
> @@ -125,427 +125,210 @@ virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
>  	}
>  }
>
> -static inline void
> -copy_virtio_net_hdr(struct virtio_net *dev, uint64_t desc_addr,
> -		    struct virtio_net_hdr_mrg_rxbuf hdr)
> +static inline uint32_t __attribute__((always_inline))
> +loop_check(struct vhost_virtqueue *vq, uint16_t avail_idx, uint32_t pkt_left)
Creating a function just for doing this doesn't make much sense.
And the function name doesn't help.
I think you should just remove this function.

>  {
> -	if (dev->vhost_hlen == sizeof(struct virtio_net_hdr_mrg_rxbuf))
> -		*(struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr = hdr;
> -	else
> -		*(struct virtio_net_hdr *)(uintptr_t)desc_addr = hdr.hdr;
> +	if (pkt_left == 0 || avail_idx == vq->last_used_idx)
> +		return 1;
> +
> +	return 0;
>  }
>
> -static inline int __attribute__((always_inline))
> -copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
> -		  struct rte_mbuf *m, uint16_t desc_idx)
> +static inline uint32_t __attribute__((always_inline))
> +enqueue_packet(struct virtio_net *dev, struct vhost_virtqueue *vq,
> +		uint16_t avail_idx, struct rte_mbuf *mbuf,
> +		uint32_t is_mrg_rxbuf)
>  {
> -	uint32_t desc_avail, desc_offset;
> -	uint32_t mbuf_avail, mbuf_offset;
> -	uint32_t cpy_len;
> +	struct virtio_net_hdr_mrg_rxbuf *virtio_hdr;
>  	struct vring_desc *desc;
> -	uint64_t desc_addr;
> -	struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
> -
> -	desc = &vq->desc[desc_idx];
> +	uint64_t desc_addr = 0;
> +	uint32_t desc_chain_head = 0;
> +	uint32_t desc_chain_len = 0;
> +	uint32_t desc_current = 0;
> +	uint32_t desc_offset = 0;
> +	uint32_t mbuf_len = 0;
> +	uint32_t mbuf_avail = 0;
> +	uint32_t copy_len = 0;
> +	uint32_t extra_buffers = 0;
> +	uint32_t used_idx_round = 0;
Most of these variables don't need to be initialized.

> +
> +	/* start with the first mbuf of the packet */
> +	mbuf_len = rte_pktmbuf_data_len(mbuf);
> +	mbuf_avail = mbuf_len;
> +
> +	/* get the current desc */
> +	desc_current = vq->avail->ring[(vq->last_used_idx) & (vq->size - 1)];
> +	desc_chain_head = desc_current;
> +	desc = &vq->desc[desc_current];
>  	desc_addr = gpa_to_vva(dev, desc->addr);
> -	/*
> -	 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
> -	 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
> -	 * otherwise stores offset on the stack instead of in a register.
> -	 */
> -	if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr)
> -		return -1;
> -
> -	rte_prefetch0((void *)(uintptr_t)desc_addr);
> +	if (unlikely(!desc_addr))
> +		goto error;
>
> -	virtio_enqueue_offload(m, &virtio_hdr.hdr);
> -	copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
> +	/* handle virtio header */
> +	virtio_hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr;
> +	virtio_enqueue_offload(mbuf, &(virtio_hdr->hdr));
Parenthesis around virtio_hdr->hdr shouldn't be needed.
>  	vhost_log_write(dev, desc->addr, dev->vhost_hlen);
> -	PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
Looks like you remove the PRINT_PACKET calls.
Does it impact performance?
In any case, it should be mentionned in the commit message.

> -
>  	desc_offset = dev->vhost_hlen;
> -	desc_avail  = desc->len - dev->vhost_hlen;
> +	desc_chain_len = desc_offset;
> +	desc_addr += desc_offset;
> +	if (is_mrg_rxbuf)
> +		virtio_hdr->num_buffers = 1;
>
> -	mbuf_avail  = rte_pktmbuf_data_len(m);
> -	mbuf_offset = 0;
> -	while (mbuf_avail != 0 || m->next != NULL) {
> -		/* done with current mbuf, fetch next */
> -		if (mbuf_avail == 0) {
> -			m = m->next;
> -
> -			mbuf_offset = 0;
> -			mbuf_avail  = rte_pktmbuf_data_len(m);
> +	/* start copy from mbuf to desc */
> +	while (1) {
Please avoid while(1) when you can check for a real condition:
while (mbuf_avail || mbuf->next) ?

Compiler should optimize this properly, no?
> +		/* get the next mbuf if the current done */
> +		if (!mbuf_avail) {
> +			if (mbuf->next) {
> +				mbuf = mbuf->next;
> +				mbuf_len = rte_pktmbuf_data_len(mbuf);
> +				mbuf_avail = mbuf_len;
> +			} else
> +				break;
>  		}
>
> -		/* done with current desc buf, fetch next */
> -		if (desc_avail == 0) {
> -			if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
> -				/* Room in vring buffer is not enough */
> -				return -1;
> -			}
> -			if (unlikely(desc->next >= vq->size))
> -				return -1;
> -
> -			desc = &vq->desc[desc->next];
> -			desc_addr = gpa_to_vva(dev, desc->addr);
> -			if (unlikely(!desc_addr))
> -				return -1;
> -
> -			desc_offset = 0;
> -			desc_avail  = desc->len;
> +		/* get the next desc if the current done */
> +		if (desc->len <= desc_offset) {
> +			if (desc->flags & VRING_DESC_F_NEXT) {
> +				/* go on with the current desc chain */
> +				desc_offset = 0;
> +				desc_current = desc->next;
> +				desc = &vq->desc[desc_current];
> +				desc_addr = gpa_to_vva(dev, desc->addr);
> +				if (unlikely(!desc_addr))
> +					goto rollback;
you could goto directly to error, and decrement last_used_idx directly
under "error"'s goto since extra_buffers will be zero otherwise.

Also, except desc_current affectation, all the above code is common
with mergeable case, so you should avoid duplication.
> +			} else if (is_mrg_rxbuf) {
> +				/* start with the next desc chain */
> +				used_idx_round = vq->last_used_idx
> +					& (vq->size - 1);
> +				vq->used->ring[used_idx_round].id =
> +					desc_chain_head;
> +				vq->used->ring[used_idx_round].len =
> +					desc_chain_len;
> +				vhost_log_used_vring(dev, vq,
> +					offsetof(struct vring_used,
> +						ring[used_idx_round]),
> +					sizeof(vq->used->ring[
> +						used_idx_round]));
> +				vq->last_used_idx++;
> +				extra_buffers++;
> +				virtio_hdr->num_buffers++;
> +				if (avail_idx == vq->last_used_idx)
> +					goto rollback;
> +
> +				desc_current =
> +					vq->avail->ring[(vq->last_used_idx) &
> +					(vq->size - 1)];
> +				desc_chain_head = desc_current;
> +				desc = &vq->desc[desc_current];
> +				desc_addr = gpa_to_vva(dev, desc->addr);
> +				if (unlikely(!desc_addr))
> +					goto rollback;
> +
> +				desc_chain_len = 0;
> +				desc_offset = 0;
> +			} else
> +				goto rollback;
>  		}
>
> -		cpy_len = RTE_MIN(desc_avail, mbuf_avail);
> -		rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
> -			rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
> -			cpy_len);
> -		vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
> -		PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
> -			     cpy_len, 0);
> -
> -		mbuf_avail  -= cpy_len;
> -		mbuf_offset += cpy_len;
> -		desc_avail  -= cpy_len;
> -		desc_offset += cpy_len;
> -	}
> -
> -	return 0;
> -}
> -
> -/**
> - * This function adds buffers to the virtio devices RX virtqueue. Buffers can
> - * be received from the physical port or from another virtio device. A packet
> - * count is returned to indicate the number of packets that are succesfully
> - * added to the RX queue. This function works when the mbuf is scattered, but
> - * it doesn't support the mergeable feature.
> - */
> -static inline uint32_t __attribute__((always_inline))
> -virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
> -	      struct rte_mbuf **pkts, uint32_t count)
> -{
> -	struct vhost_virtqueue *vq;
> -	uint16_t avail_idx, free_entries, start_idx;
> -	uint16_t desc_indexes[MAX_PKT_BURST];
> -	uint16_t used_idx;
> -	uint32_t i;
> -
> -	LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
> -	if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
> -		RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
> -			dev->vid, __func__, queue_id);
> -		return 0;
> +		/* copy mbuf data */
> +		copy_len = RTE_MIN(desc->len - desc_offset, mbuf_avail);
> +		rte_memcpy((void *)(uintptr_t)desc_addr,
> +				rte_pktmbuf_mtod_offset(mbuf, void *,
> +					mbuf_len - mbuf_avail),
> +				copy_len);
> +		vhost_log_write(dev, desc->addr + desc_offset, copy_len);
> +		mbuf_avail -= copy_len;
> +		desc_offset += copy_len;
> +		desc_addr += copy_len;
> +		desc_chain_len += copy_len;
>  	}
>
> -	vq = dev->virtqueue[queue_id];
> -	if (unlikely(vq->enabled == 0))
> -		return 0;
> -
> -	avail_idx = *((volatile uint16_t *)&vq->avail->idx);
> -	start_idx = vq->last_used_idx;
> -	free_entries = avail_idx - start_idx;
> -	count = RTE_MIN(count, free_entries);
> -	count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
> -	if (count == 0)
> -		return 0;
> -
> -	LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
> -		dev->vid, start_idx, start_idx + count);
> -
> -	/* Retrieve all of the desc indexes first to avoid caching issues. */
> -	rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
> -	for (i = 0; i < count; i++) {
> -		used_idx = (start_idx + i) & (vq->size - 1);
> -		desc_indexes[i] = vq->avail->ring[used_idx];
> -		vq->used->ring[used_idx].id = desc_indexes[i];
> -		vq->used->ring[used_idx].len = pkts[i]->pkt_len +
> -					       dev->vhost_hlen;
> -		vhost_log_used_vring(dev, vq,
> -			offsetof(struct vring_used, ring[used_idx]),
> -			sizeof(vq->used->ring[used_idx]));
> -	}
> +	used_idx_round = vq->last_used_idx & (vq->size - 1);
> +	vq->used->ring[used_idx_round].id = desc_chain_head;
> +	vq->used->ring[used_idx_round].len = desc_chain_len;
> +	vhost_log_used_vring(dev, vq,
> +		offsetof(struct vring_used, ring[used_idx_round]),
> +		sizeof(vq->used->ring[used_idx_round]));
> +	vq->last_used_idx++;
All this code is duplicatedd from the rx_mergeable base.
I think a dedicated inline function would really make sense here.

>
> -	rte_prefetch0(&vq->desc[desc_indexes[0]]);
> -	for (i = 0; i < count; i++) {
> -		uint16_t desc_idx = desc_indexes[i];
> -		int err;
> +	return 0;
>
> -		err = copy_mbuf_to_desc(dev, vq, pkts[i], desc_idx);
> -		if (unlikely(err)) {
> -			used_idx = (start_idx + i) & (vq->size - 1);
> -			vq->used->ring[used_idx].len = dev->vhost_hlen;
> -			vhost_log_used_vring(dev, vq,
> -				offsetof(struct vring_used, ring[used_idx]),
> -				sizeof(vq->used->ring[used_idx]));
> -		}
> +rollback:
> +	/* rollback on any error if last_used_idx update on-the-fly */
> +	if (is_mrg_rxbuf)
If (!is_mrg_rxbuf), extra_buffers will be zero, so just remove the test,
and place belw line directly under error: as explained above.
> +		vq->last_used_idx -= extra_buffers;
>
> -		if (i + 1 < count)
> -			rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
> -	}
> +error:
> +	return 1;
> +}
>
> +static inline void __attribute__((always_inline))
> +notify_guest(struct virtio_net *dev, struct vhost_virtqueue *vq)
> +{
>  	rte_smp_wmb();
> -
> -	*(volatile uint16_t *)&vq->used->idx += count;
> -	vq->last_used_idx += count;
> -	vhost_log_used_vring(dev, vq,
> -		offsetof(struct vring_used, idx),
> -		sizeof(vq->used->idx));
> -
> -	/* flush used->idx update before we read avail->flags. */
> +	vq->used->idx = vq->last_used_idx;
> +	vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
> +			sizeof(vq->used->idx));
>  	rte_mb();
> -
> -	/* Kick the guest if necessary. */
>  	if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
>  			&& (vq->callfd >= 0))
>  		eventfd_write(vq->callfd, (eventfd_t)1);
> -	return count;
> -}
> -
> -static inline int
> -fill_vec_buf(struct vhost_virtqueue *vq, uint32_t avail_idx,
> -	     uint32_t *allocated, uint32_t *vec_idx,
> -	     struct buf_vector *buf_vec)
> -{
> -	uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
> -	uint32_t vec_id = *vec_idx;
> -	uint32_t len    = *allocated;
> -
> -	while (1) {
> -		if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
> -			return -1;
> -
> -		len += vq->desc[idx].len;
> -		buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
> -		buf_vec[vec_id].buf_len  = vq->desc[idx].len;
> -		buf_vec[vec_id].desc_idx = idx;
> -		vec_id++;
> -
> -		if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
> -			break;
> -
> -		idx = vq->desc[idx].next;
> -	}
> -
> -	*allocated = len;
> -	*vec_idx   = vec_id;
> -
> -	return 0;
> -}
> -
> -/*
> - * Returns -1 on fail, 0 on success
> - */
> -static inline int
> -reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
> -			    uint16_t *end, struct buf_vector *buf_vec)
> -{
> -	uint16_t cur_idx;
> -	uint16_t avail_idx;
> -	uint32_t allocated = 0;
> -	uint32_t vec_idx = 0;
> -	uint16_t tries = 0;
> -
> -	cur_idx  = vq->last_used_idx;
> -
> -	while (1) {
> -		avail_idx = *((volatile uint16_t *)&vq->avail->idx);
> -		if (unlikely(cur_idx == avail_idx))
> -			return -1;
> -
> -		if (unlikely(fill_vec_buf(vq, cur_idx, &allocated,
> -					  &vec_idx, buf_vec) < 0))
> -			return -1;
> -
> -		cur_idx++;
> -		tries++;
> -
> -		if (allocated >= size)
> -			break;
> -
> -		/*
> -		 * if we tried all available ring items, and still
> -		 * can't get enough buf, it means something abnormal
> -		 * happened.
> -		 */
> -		if (unlikely(tries >= vq->size))
> -			return -1;
> -	}
> -
> -	*end = cur_idx;
> -	return 0;
>  }
>
> -static inline uint32_t __attribute__((always_inline))
> -copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
> -			    uint16_t end_idx, struct rte_mbuf *m,
> -			    struct buf_vector *buf_vec)
> +uint16_t
> +rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
> +	struct rte_mbuf **pkts, uint16_t count)
>  {
> -	struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
> -	uint32_t vec_idx = 0;
> -	uint16_t start_idx = vq->last_used_idx;
> -	uint16_t cur_idx = start_idx;
> -	uint64_t desc_addr;
> -	uint32_t mbuf_offset, mbuf_avail;
> -	uint32_t desc_offset, desc_avail;
> -	uint32_t cpy_len;
> -	uint16_t desc_idx, used_idx;
> -
> -	if (unlikely(m == NULL))
> +	struct vhost_virtqueue *vq;
> +	struct virtio_net *dev;
> +	uint32_t pkt_idx = 0;
> +	uint32_t pkt_left = 0;
> +	uint32_t pkt_sent = 0;
> +	uint32_t is_mrg_rxbuf = 0;
> +	uint16_t avail_idx = 0;
> +
> +	/* precheck */
Comment not very informative here.
> +	if (unlikely(count == 0))
>  		return 0;
>
> -	LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
> -		dev->vid, cur_idx, end_idx);
> +	count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
>
> -	desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
> -	if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr)
> +	dev = get_device(vid);
> +	if (unlikely(!dev))
>  		return 0;
>
> -	rte_prefetch0((void *)(uintptr_t)desc_addr);
> -
> -	virtio_hdr.num_buffers = end_idx - start_idx;
> -	LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
> -		dev->vid, virtio_hdr.num_buffers);
> -
> -	virtio_enqueue_offload(m, &virtio_hdr.hdr);
> -	copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
> -	vhost_log_write(dev, buf_vec[vec_idx].buf_addr, dev->vhost_hlen);
> -	PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
> -
> -	desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
> -	desc_offset = dev->vhost_hlen;
> -
> -	mbuf_avail  = rte_pktmbuf_data_len(m);
> -	mbuf_offset = 0;
> -	while (mbuf_avail != 0 || m->next != NULL) {
> -		/* done with current desc buf, get the next one */
> -		if (desc_avail == 0) {
> -			desc_idx = buf_vec[vec_idx].desc_idx;
> -
> -			if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
> -				/* Update used ring with desc information */
> -				used_idx = cur_idx++ & (vq->size - 1);
> -				vq->used->ring[used_idx].id  = desc_idx;
> -				vq->used->ring[used_idx].len = desc_offset;
> -				vhost_log_used_vring(dev, vq,
> -					offsetof(struct vring_used,
> -						 ring[used_idx]),
> -					sizeof(vq->used->ring[used_idx]));
> -			}
> -
> -			vec_idx++;
> -			desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
> -			if (unlikely(!desc_addr))
> -				return 0;
> -
> -			/* Prefetch buffer address. */
> -			rte_prefetch0((void *)(uintptr_t)desc_addr);
> -			desc_offset = 0;
> -			desc_avail  = buf_vec[vec_idx].buf_len;
> -		}
> -
> -		/* done with current mbuf, get the next one */
> -		if (mbuf_avail == 0) {
> -			m = m->next;
> -
> -			mbuf_offset = 0;
> -			mbuf_avail  = rte_pktmbuf_data_len(m);
> -		}
> -
> -		cpy_len = RTE_MIN(desc_avail, mbuf_avail);
> -		rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
> -			rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
> -			cpy_len);
> -		vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
> -			cpy_len);
> -		PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
> -			cpy_len, 0);
> -
> -		mbuf_avail  -= cpy_len;
> -		mbuf_offset += cpy_len;
> -		desc_avail  -= cpy_len;
> -		desc_offset += cpy_len;
> -	}
> -
> -	used_idx = cur_idx & (vq->size - 1);
> -	vq->used->ring[used_idx].id = buf_vec[vec_idx].desc_idx;
> -	vq->used->ring[used_idx].len = desc_offset;
> -	vhost_log_used_vring(dev, vq,
> -		offsetof(struct vring_used, ring[used_idx]),
> -		sizeof(vq->used->ring[used_idx]));
> -
> -	return end_idx - start_idx;
> -}
> -
> -static inline uint32_t __attribute__((always_inline))
> -virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
> -	struct rte_mbuf **pkts, uint32_t count)
> -{
> -	struct vhost_virtqueue *vq;
> -	uint32_t pkt_idx = 0, nr_used = 0;
> -	uint16_t end;
> -	struct buf_vector buf_vec[BUF_VECTOR_MAX];
> -
> -	LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
> -	if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
> -		RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
> -			dev->vid, __func__, queue_id);
> +	if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb)))
>  		return 0;
> -	}
>
>  	vq = dev->virtqueue[queue_id];
> -	if (unlikely(vq->enabled == 0))
> +	if (unlikely(!vq->enabled))
>  		return 0;
>
> -	count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
> -	if (count == 0)
> -		return 0;
> +	if (dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF))
> +		is_mrg_rxbuf = 1;
>
> -	for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
> -		uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
> -
> -		if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len,
> -							 &end, buf_vec) < 0)) {
> -			LOG_DEBUG(VHOST_DATA,
> -				"(%d) failed to get enough desc from vring\n",
> -				dev->vid);
> +	/* start enqueuing packets 1 by 1 */
> +	pkt_idx = 0;
> +	pkt_left = count;
> +	avail_idx = *((volatile uint16_t *)&vq->avail->idx);
> +	while (1) {
> +		if (loop_check(vq, avail_idx, pkt_left))
What about:
while (pkt_left && avail_idx != vq->last_used_idx) {

>  			break;
> -		}
> -
> -		nr_used = copy_mbuf_to_desc_mergeable(dev, vq, end,
> -						      pkts[pkt_idx], buf_vec);
> -		rte_smp_wmb();
>
> -		*(volatile uint16_t *)&vq->used->idx += nr_used;
> -		vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
> -			sizeof(vq->used->idx));
> -		vq->last_used_idx += nr_used;
> -	}
> -
> -	if (likely(pkt_idx)) {
> -		/* flush used->idx update before we read avail->flags. */
> -		rte_mb();
> +		if (enqueue_packet(dev, vq, avail_idx, pkts[pkt_idx],
> +					is_mrg_rxbuf))
> +			break;
>
> -		/* Kick the guest if necessary. */
> -		if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
> -				&& (vq->callfd >= 0))
> -			eventfd_write(vq->callfd, (eventfd_t)1);
> +		pkt_idx++;
> +		pkt_sent++;
> +		pkt_left--;
>  	}
>
> -	return pkt_idx;
> -}
> -
> -uint16_t
> -rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
> -	struct rte_mbuf **pkts, uint16_t count)
> -{
> -	struct virtio_net *dev = get_device(vid);
> -
> -	if (!dev)
> -		return 0;
> +	/* update used idx and kick the guest if necessary */
> +	if (pkt_sent)
> +		notify_guest(dev, vq);
>
> -	if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
> -		return virtio_dev_merge_rx(dev, queue_id, pkts, count);
> -	else
> -		return virtio_dev_rx(dev, queue_id, pkts, count);
> +	return pkt_sent;
>  }
>
>  static void
>
Zhihong Wang Aug. 23, 2016, 2:27 a.m. UTC | #2
Hi Maxime,

Thanks very much for the detailed review.

> -----Original Message-----
> From: Maxime Coquelin [mailto:maxime.coquelin@redhat.com]
> Sent: Monday, August 22, 2016 5:36 PM
> To: Wang, Zhihong <zhihong.wang@intel.com>; dev@dpdk.org
> Cc: yuanhan.liu@linux.intel.com
> Subject: Re: [PATCH v3 1/5] vhost: rewrite enqueue
> 
> 
> 
> On 08/19/2016 07:43 AM, Zhihong Wang wrote:
> > This patch implements the vhost logic from scratch into a single function
> > designed for high performance and better maintainability.
> >
> > ---
> > Changes in v3:
> >
> >  1. Rewrite enqueue and delete the obsolete in the same patch.
> >
> > Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
> > ---
> >  lib/librte_vhost/vhost_rxtx.c | 537 +++++++++++++-----------------------------
> >  1 file changed, 160 insertions(+), 377 deletions(-)
> >
> > diff --git a/lib/librte_vhost/vhost_rxtx.c b/lib/librte_vhost/vhost_rxtx.c
> > index 08a73fd..b09a9c3 100644
> > --- a/lib/librte_vhost/vhost_rxtx.c
> > +++ b/lib/librte_vhost/vhost_rxtx.c
> > @@ -91,7 +91,7 @@ is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t
> qp_nb)
> >  	return (is_tx ^ (idx & 1)) == 0 && idx < qp_nb * VIRTIO_QNUM;
> >  }
> >
> > -static void
> > +static inline void __attribute__((always_inline))
> >  virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr
> *net_hdr)
> >  {
> >  	if (m_buf->ol_flags & PKT_TX_L4_MASK) {
> > @@ -125,427 +125,210 @@ virtio_enqueue_offload(struct rte_mbuf *m_buf,
> struct virtio_net_hdr *net_hdr)
> >  	}
> >  }
> >
> > -static inline void
> > -copy_virtio_net_hdr(struct virtio_net *dev, uint64_t desc_addr,
> > -		    struct virtio_net_hdr_mrg_rxbuf hdr)
> > +static inline uint32_t __attribute__((always_inline))
> > +loop_check(struct vhost_virtqueue *vq, uint16_t avail_idx, uint32_t pkt_left)
> Creating a function just for doing this doesn't make much sense.
> And the function name doesn't help.
> I think you should just remove this function.

Okay.

> 
> >  {
> > -	if (dev->vhost_hlen == sizeof(struct virtio_net_hdr_mrg_rxbuf))
> > -		*(struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr = hdr;
> > -	else
> > -		*(struct virtio_net_hdr *)(uintptr_t)desc_addr = hdr.hdr;
> > +	if (pkt_left == 0 || avail_idx == vq->last_used_idx)
> > +		return 1;
> > +
> > +	return 0;
> >  }
> >
> > -static inline int __attribute__((always_inline))
> > -copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
> > -		  struct rte_mbuf *m, uint16_t desc_idx)
> > +static inline uint32_t __attribute__((always_inline))
> > +enqueue_packet(struct virtio_net *dev, struct vhost_virtqueue *vq,
> > +		uint16_t avail_idx, struct rte_mbuf *mbuf,
> > +		uint32_t is_mrg_rxbuf)
> >  {
> > -	uint32_t desc_avail, desc_offset;
> > -	uint32_t mbuf_avail, mbuf_offset;
> > -	uint32_t cpy_len;
> > +	struct virtio_net_hdr_mrg_rxbuf *virtio_hdr;
> >  	struct vring_desc *desc;
> > -	uint64_t desc_addr;
> > -	struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
> > -
> > -	desc = &vq->desc[desc_idx];
> > +	uint64_t desc_addr = 0;
> > +	uint32_t desc_chain_head = 0;
> > +	uint32_t desc_chain_len = 0;
> > +	uint32_t desc_current = 0;
> > +	uint32_t desc_offset = 0;
> > +	uint32_t mbuf_len = 0;
> > +	uint32_t mbuf_avail = 0;
> > +	uint32_t copy_len = 0;
> > +	uint32_t extra_buffers = 0;
> > +	uint32_t used_idx_round = 0;
> Most of these variables don't need to be initialized.

Okay.

> 
> > +
> > +	/* start with the first mbuf of the packet */
> > +	mbuf_len = rte_pktmbuf_data_len(mbuf);
> > +	mbuf_avail = mbuf_len;
> > +
> > +	/* get the current desc */
> > +	desc_current = vq->avail->ring[(vq->last_used_idx) & (vq->size - 1)];
> > +	desc_chain_head = desc_current;
> > +	desc = &vq->desc[desc_current];
> >  	desc_addr = gpa_to_vva(dev, desc->addr);
> > -	/*
> > -	 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
> > -	 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
> > -	 * otherwise stores offset on the stack instead of in a register.
> > -	 */
> > -	if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr)
> > -		return -1;
> > -
> > -	rte_prefetch0((void *)(uintptr_t)desc_addr);
> > +	if (unlikely(!desc_addr))
> > +		goto error;
> >
> > -	virtio_enqueue_offload(m, &virtio_hdr.hdr);
> > -	copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
> > +	/* handle virtio header */
> > +	virtio_hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr;
> > +	virtio_enqueue_offload(mbuf, &(virtio_hdr->hdr));
> Parenthesis around virtio_hdr->hdr shouldn't be needed.
> >  	vhost_log_write(dev, desc->addr, dev->vhost_hlen);
> > -	PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
> Looks like you remove the PRINT_PACKET calls.
> Does it impact performance?
> In any case, it should be mentionned in the commit message.

Will add this.

> 
> > -
> >  	desc_offset = dev->vhost_hlen;
> > -	desc_avail  = desc->len - dev->vhost_hlen;
> > +	desc_chain_len = desc_offset;
> > +	desc_addr += desc_offset;
> > +	if (is_mrg_rxbuf)
> > +		virtio_hdr->num_buffers = 1;
> >
> > -	mbuf_avail  = rte_pktmbuf_data_len(m);
> > -	mbuf_offset = 0;
> > -	while (mbuf_avail != 0 || m->next != NULL) {
> > -		/* done with current mbuf, fetch next */
> > -		if (mbuf_avail == 0) {
> > -			m = m->next;
> > -
> > -			mbuf_offset = 0;
> > -			mbuf_avail  = rte_pktmbuf_data_len(m);
> > +	/* start copy from mbuf to desc */
> > +	while (1) {
> Please avoid while(1) when you can check for a real condition:
> while (mbuf_avail || mbuf->next) ?

Will rewrite this logic.

> 
> Compiler should optimize this properly, no?
> > +		/* get the next mbuf if the current done */
> > +		if (!mbuf_avail) {
> > +			if (mbuf->next) {
> > +				mbuf = mbuf->next;
> > +				mbuf_len = rte_pktmbuf_data_len(mbuf);
> > +				mbuf_avail = mbuf_len;
> > +			} else
> > +				break;
> >  		}
> >
> > -		/* done with current desc buf, fetch next */
> > -		if (desc_avail == 0) {
> > -			if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
> > -				/* Room in vring buffer is not enough */
> > -				return -1;
> > -			}
> > -			if (unlikely(desc->next >= vq->size))
> > -				return -1;
> > -
> > -			desc = &vq->desc[desc->next];
> > -			desc_addr = gpa_to_vva(dev, desc->addr);
> > -			if (unlikely(!desc_addr))
> > -				return -1;
> > -
> > -			desc_offset = 0;
> > -			desc_avail  = desc->len;
> > +		/* get the next desc if the current done */
> > +		if (desc->len <= desc_offset) {
> > +			if (desc->flags & VRING_DESC_F_NEXT) {
> > +				/* go on with the current desc chain */
> > +				desc_offset = 0;
> > +				desc_current = desc->next;
> > +				desc = &vq->desc[desc_current];
> > +				desc_addr = gpa_to_vva(dev, desc->addr);
> > +				if (unlikely(!desc_addr))
> > +					goto rollback;
> you could goto directly to error, and decrement last_used_idx directly
> under "error"'s goto since extra_buffers will be zero otherwise.

Good call.

> 
> Also, except desc_current affectation, all the above code is common
> with mergeable case, so you should avoid duplication.
> > +			} else if (is_mrg_rxbuf) {
> > +				/* start with the next desc chain */
> > +				used_idx_round = vq->last_used_idx
> > +					& (vq->size - 1);
> > +				vq->used->ring[used_idx_round].id =
> > +					desc_chain_head;
> > +				vq->used->ring[used_idx_round].len =
> > +					desc_chain_len;
> > +				vhost_log_used_vring(dev, vq,
> > +					offsetof(struct vring_used,
> > +						ring[used_idx_round]),
> > +					sizeof(vq->used->ring[
> > +						used_idx_round]));
> > +				vq->last_used_idx++;
> > +				extra_buffers++;
> > +				virtio_hdr->num_buffers++;
> > +				if (avail_idx == vq->last_used_idx)
> > +					goto rollback;
> > +
> > +				desc_current =
> > +					vq->avail->ring[(vq->last_used_idx) &
> > +					(vq->size - 1)];
> > +				desc_chain_head = desc_current;
> > +				desc = &vq->desc[desc_current];
> > +				desc_addr = gpa_to_vva(dev, desc->addr);
> > +				if (unlikely(!desc_addr))
> > +					goto rollback;
> > +
> > +				desc_chain_len = 0;
> > +				desc_offset = 0;
> > +			} else
> > +				goto rollback;
> >  		}
> >
> > -		cpy_len = RTE_MIN(desc_avail, mbuf_avail);
> > -		rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
> > -			rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
> > -			cpy_len);
> > -		vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
> > -		PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
> > -			     cpy_len, 0);
> > -
> > -		mbuf_avail  -= cpy_len;
> > -		mbuf_offset += cpy_len;
> > -		desc_avail  -= cpy_len;
> > -		desc_offset += cpy_len;
> > -	}
> > -
> > -	return 0;
> > -}
> > -
> > -/**
> > - * This function adds buffers to the virtio devices RX virtqueue. Buffers can
> > - * be received from the physical port or from another virtio device. A packet
> > - * count is returned to indicate the number of packets that are succesfully
> > - * added to the RX queue. This function works when the mbuf is scattered, but
> > - * it doesn't support the mergeable feature.
> > - */
> > -static inline uint32_t __attribute__((always_inline))
> > -virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
> > -	      struct rte_mbuf **pkts, uint32_t count)
> > -{
> > -	struct vhost_virtqueue *vq;
> > -	uint16_t avail_idx, free_entries, start_idx;
> > -	uint16_t desc_indexes[MAX_PKT_BURST];
> > -	uint16_t used_idx;
> > -	uint32_t i;
> > -
> > -	LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
> > -	if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
> > -		RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
> > -			dev->vid, __func__, queue_id);
> > -		return 0;
> > +		/* copy mbuf data */
> > +		copy_len = RTE_MIN(desc->len - desc_offset, mbuf_avail);
> > +		rte_memcpy((void *)(uintptr_t)desc_addr,
> > +				rte_pktmbuf_mtod_offset(mbuf, void *,
> > +					mbuf_len - mbuf_avail),
> > +				copy_len);
> > +		vhost_log_write(dev, desc->addr + desc_offset, copy_len);
> > +		mbuf_avail -= copy_len;
> > +		desc_offset += copy_len;
> > +		desc_addr += copy_len;
> > +		desc_chain_len += copy_len;
> >  	}
> >
> > -	vq = dev->virtqueue[queue_id];
> > -	if (unlikely(vq->enabled == 0))
> > -		return 0;
> > -
> > -	avail_idx = *((volatile uint16_t *)&vq->avail->idx);
> > -	start_idx = vq->last_used_idx;
> > -	free_entries = avail_idx - start_idx;
> > -	count = RTE_MIN(count, free_entries);
> > -	count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
> > -	if (count == 0)
> > -		return 0;
> > -
> > -	LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
> > -		dev->vid, start_idx, start_idx + count);
> > -
> > -	/* Retrieve all of the desc indexes first to avoid caching issues. */
> > -	rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
> > -	for (i = 0; i < count; i++) {
> > -		used_idx = (start_idx + i) & (vq->size - 1);
> > -		desc_indexes[i] = vq->avail->ring[used_idx];
> > -		vq->used->ring[used_idx].id = desc_indexes[i];
> > -		vq->used->ring[used_idx].len = pkts[i]->pkt_len +
> > -					       dev->vhost_hlen;
> > -		vhost_log_used_vring(dev, vq,
> > -			offsetof(struct vring_used, ring[used_idx]),
> > -			sizeof(vq->used->ring[used_idx]));
> > -	}
> > +	used_idx_round = vq->last_used_idx & (vq->size - 1);
> > +	vq->used->ring[used_idx_round].id = desc_chain_head;
> > +	vq->used->ring[used_idx_round].len = desc_chain_len;
> > +	vhost_log_used_vring(dev, vq,
> > +		offsetof(struct vring_used, ring[used_idx_round]),
> > +		sizeof(vq->used->ring[used_idx_round]));
> > +	vq->last_used_idx++;
> All this code is duplicatedd from the rx_mergeable base.
> I think a dedicated inline function would really make sense here.

Good catch. Will make a function for this.

> 
> >
> > -	rte_prefetch0(&vq->desc[desc_indexes[0]]);
> > -	for (i = 0; i < count; i++) {
> > -		uint16_t desc_idx = desc_indexes[i];
> > -		int err;
> > +	return 0;
> >
> > -		err = copy_mbuf_to_desc(dev, vq, pkts[i], desc_idx);
> > -		if (unlikely(err)) {
> > -			used_idx = (start_idx + i) & (vq->size - 1);
> > -			vq->used->ring[used_idx].len = dev->vhost_hlen;
> > -			vhost_log_used_vring(dev, vq,
> > -				offsetof(struct vring_used, ring[used_idx]),
> > -				sizeof(vq->used->ring[used_idx]));
> > -		}
> > +rollback:
> > +	/* rollback on any error if last_used_idx update on-the-fly */
> > +	if (is_mrg_rxbuf)
> If (!is_mrg_rxbuf), extra_buffers will be zero, so just remove the test,
> and place belw line directly under error: as explained above.

Sure. Thanks.

> > +		vq->last_used_idx -= extra_buffers;
> >
> > -		if (i + 1 < count)
> > -			rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
> > -	}
> > +error:
> > +	return 1;
> > +}
> >
> > +static inline void __attribute__((always_inline))
> > +notify_guest(struct virtio_net *dev, struct vhost_virtqueue *vq)
> > +{
> >  	rte_smp_wmb();
> > -
> > -	*(volatile uint16_t *)&vq->used->idx += count;
> > -	vq->last_used_idx += count;
> > -	vhost_log_used_vring(dev, vq,
> > -		offsetof(struct vring_used, idx),
> > -		sizeof(vq->used->idx));
> > -
> > -	/* flush used->idx update before we read avail->flags. */
> > +	vq->used->idx = vq->last_used_idx;
> > +	vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
> > +			sizeof(vq->used->idx));
> >  	rte_mb();
> > -
> > -	/* Kick the guest if necessary. */
> >  	if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
> >  			&& (vq->callfd >= 0))
> >  		eventfd_write(vq->callfd, (eventfd_t)1);
> > -	return count;
> > -}
> > -
> > -static inline int
> > -fill_vec_buf(struct vhost_virtqueue *vq, uint32_t avail_idx,
> > -	     uint32_t *allocated, uint32_t *vec_idx,
> > -	     struct buf_vector *buf_vec)
> > -{
> > -	uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
> > -	uint32_t vec_id = *vec_idx;
> > -	uint32_t len    = *allocated;
> > -
> > -	while (1) {
> > -		if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
> > -			return -1;
> > -
> > -		len += vq->desc[idx].len;
> > -		buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
> > -		buf_vec[vec_id].buf_len  = vq->desc[idx].len;
> > -		buf_vec[vec_id].desc_idx = idx;
> > -		vec_id++;
> > -
> > -		if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
> > -			break;
> > -
> > -		idx = vq->desc[idx].next;
> > -	}
> > -
> > -	*allocated = len;
> > -	*vec_idx   = vec_id;
> > -
> > -	return 0;
> > -}
> > -
> > -/*
> > - * Returns -1 on fail, 0 on success
> > - */
> > -static inline int
> > -reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
> > -			    uint16_t *end, struct buf_vector *buf_vec)
> > -{
> > -	uint16_t cur_idx;
> > -	uint16_t avail_idx;
> > -	uint32_t allocated = 0;
> > -	uint32_t vec_idx = 0;
> > -	uint16_t tries = 0;
> > -
> > -	cur_idx  = vq->last_used_idx;
> > -
> > -	while (1) {
> > -		avail_idx = *((volatile uint16_t *)&vq->avail->idx);
> > -		if (unlikely(cur_idx == avail_idx))
> > -			return -1;
> > -
> > -		if (unlikely(fill_vec_buf(vq, cur_idx, &allocated,
> > -					  &vec_idx, buf_vec) < 0))
> > -			return -1;
> > -
> > -		cur_idx++;
> > -		tries++;
> > -
> > -		if (allocated >= size)
> > -			break;
> > -
> > -		/*
> > -		 * if we tried all available ring items, and still
> > -		 * can't get enough buf, it means something abnormal
> > -		 * happened.
> > -		 */
> > -		if (unlikely(tries >= vq->size))
> > -			return -1;
> > -	}
> > -
> > -	*end = cur_idx;
> > -	return 0;
> >  }
> >
> > -static inline uint32_t __attribute__((always_inline))
> > -copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct
> vhost_virtqueue *vq,
> > -			    uint16_t end_idx, struct rte_mbuf *m,
> > -			    struct buf_vector *buf_vec)
> > +uint16_t
> > +rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
> > +	struct rte_mbuf **pkts, uint16_t count)
> >  {
> > -	struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
> > -	uint32_t vec_idx = 0;
> > -	uint16_t start_idx = vq->last_used_idx;
> > -	uint16_t cur_idx = start_idx;
> > -	uint64_t desc_addr;
> > -	uint32_t mbuf_offset, mbuf_avail;
> > -	uint32_t desc_offset, desc_avail;
> > -	uint32_t cpy_len;
> > -	uint16_t desc_idx, used_idx;
> > -
> > -	if (unlikely(m == NULL))
> > +	struct vhost_virtqueue *vq;
> > +	struct virtio_net *dev;
> > +	uint32_t pkt_idx = 0;
> > +	uint32_t pkt_left = 0;
> > +	uint32_t pkt_sent = 0;
> > +	uint32_t is_mrg_rxbuf = 0;
> > +	uint16_t avail_idx = 0;
> > +
> > +	/* precheck */
> Comment not very informative here.

Okay.

> > +	if (unlikely(count == 0))
> >  		return 0;
> >
> > -	LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
> > -		dev->vid, cur_idx, end_idx);
> > +	count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
> >
> > -	desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
> > -	if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr)
> > +	dev = get_device(vid);
> > +	if (unlikely(!dev))
> >  		return 0;
> >
> > -	rte_prefetch0((void *)(uintptr_t)desc_addr);
> > -
> > -	virtio_hdr.num_buffers = end_idx - start_idx;
> > -	LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
> > -		dev->vid, virtio_hdr.num_buffers);
> > -
> > -	virtio_enqueue_offload(m, &virtio_hdr.hdr);
> > -	copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
> > -	vhost_log_write(dev, buf_vec[vec_idx].buf_addr, dev->vhost_hlen);
> > -	PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
> > -
> > -	desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
> > -	desc_offset = dev->vhost_hlen;
> > -
> > -	mbuf_avail  = rte_pktmbuf_data_len(m);
> > -	mbuf_offset = 0;
> > -	while (mbuf_avail != 0 || m->next != NULL) {
> > -		/* done with current desc buf, get the next one */
> > -		if (desc_avail == 0) {
> > -			desc_idx = buf_vec[vec_idx].desc_idx;
> > -
> > -			if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
> > -				/* Update used ring with desc information */
> > -				used_idx = cur_idx++ & (vq->size - 1);
> > -				vq->used->ring[used_idx].id  = desc_idx;
> > -				vq->used->ring[used_idx].len = desc_offset;
> > -				vhost_log_used_vring(dev, vq,
> > -					offsetof(struct vring_used,
> > -						 ring[used_idx]),
> > -					sizeof(vq->used->ring[used_idx]));
> > -			}
> > -
> > -			vec_idx++;
> > -			desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
> > -			if (unlikely(!desc_addr))
> > -				return 0;
> > -
> > -			/* Prefetch buffer address. */
> > -			rte_prefetch0((void *)(uintptr_t)desc_addr);
> > -			desc_offset = 0;
> > -			desc_avail  = buf_vec[vec_idx].buf_len;
> > -		}
> > -
> > -		/* done with current mbuf, get the next one */
> > -		if (mbuf_avail == 0) {
> > -			m = m->next;
> > -
> > -			mbuf_offset = 0;
> > -			mbuf_avail  = rte_pktmbuf_data_len(m);
> > -		}
> > -
> > -		cpy_len = RTE_MIN(desc_avail, mbuf_avail);
> > -		rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
> > -			rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
> > -			cpy_len);
> > -		vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
> > -			cpy_len);
> > -		PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
> > -			cpy_len, 0);
> > -
> > -		mbuf_avail  -= cpy_len;
> > -		mbuf_offset += cpy_len;
> > -		desc_avail  -= cpy_len;
> > -		desc_offset += cpy_len;
> > -	}
> > -
> > -	used_idx = cur_idx & (vq->size - 1);
> > -	vq->used->ring[used_idx].id = buf_vec[vec_idx].desc_idx;
> > -	vq->used->ring[used_idx].len = desc_offset;
> > -	vhost_log_used_vring(dev, vq,
> > -		offsetof(struct vring_used, ring[used_idx]),
> > -		sizeof(vq->used->ring[used_idx]));
> > -
> > -	return end_idx - start_idx;
> > -}
> > -
> > -static inline uint32_t __attribute__((always_inline))
> > -virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
> > -	struct rte_mbuf **pkts, uint32_t count)
> > -{
> > -	struct vhost_virtqueue *vq;
> > -	uint32_t pkt_idx = 0, nr_used = 0;
> > -	uint16_t end;
> > -	struct buf_vector buf_vec[BUF_VECTOR_MAX];
> > -
> > -	LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
> > -	if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
> > -		RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
> > -			dev->vid, __func__, queue_id);
> > +	if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb)))
> >  		return 0;
> > -	}
> >
> >  	vq = dev->virtqueue[queue_id];
> > -	if (unlikely(vq->enabled == 0))
> > +	if (unlikely(!vq->enabled))
> >  		return 0;
> >
> > -	count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
> > -	if (count == 0)
> > -		return 0;
> > +	if (dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF))
> > +		is_mrg_rxbuf = 1;
> >
> > -	for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
> > -		uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
> > -
> > -		if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len,
> > -							 &end, buf_vec) < 0)) {
> > -			LOG_DEBUG(VHOST_DATA,
> > -				"(%d) failed to get enough desc from vring\n",
> > -				dev->vid);
> > +	/* start enqueuing packets 1 by 1 */
> > +	pkt_idx = 0;
> > +	pkt_left = count;
> > +	avail_idx = *((volatile uint16_t *)&vq->avail->idx);
> > +	while (1) {
> > +		if (loop_check(vq, avail_idx, pkt_left))
> What about:
> while (pkt_left && avail_idx != vq->last_used_idx) {

Will rewrite it.

> 
> >  			break;
> > -		}
> > -
> > -		nr_used = copy_mbuf_to_desc_mergeable(dev, vq, end,
> > -						      pkts[pkt_idx], buf_vec);
> > -		rte_smp_wmb();
> >
> > -		*(volatile uint16_t *)&vq->used->idx += nr_used;
> > -		vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
> > -			sizeof(vq->used->idx));
> > -		vq->last_used_idx += nr_used;
> > -	}
> > -
> > -	if (likely(pkt_idx)) {
> > -		/* flush used->idx update before we read avail->flags. */
> > -		rte_mb();
> > +		if (enqueue_packet(dev, vq, avail_idx, pkts[pkt_idx],
> > +					is_mrg_rxbuf))
> > +			break;
> >
> > -		/* Kick the guest if necessary. */
> > -		if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
> > -				&& (vq->callfd >= 0))
> > -			eventfd_write(vq->callfd, (eventfd_t)1);
> > +		pkt_idx++;
> > +		pkt_sent++;
> > +		pkt_left--;
> >  	}
> >
> > -	return pkt_idx;
> > -}
> > -
> > -uint16_t
> > -rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
> > -	struct rte_mbuf **pkts, uint16_t count)
> > -{
> > -	struct virtio_net *dev = get_device(vid);
> > -
> > -	if (!dev)
> > -		return 0;
> > +	/* update used idx and kick the guest if necessary */
> > +	if (pkt_sent)
> > +		notify_guest(dev, vq);
> >
> > -	if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
> > -		return virtio_dev_merge_rx(dev, queue_id, pkts, count);
> > -	else
> > -		return virtio_dev_rx(dev, queue_id, pkts, count);
> > +	return pkt_sent;
> >  }
> >
> >  static void
> >
Yuanhan Liu Aug. 25, 2016, 4 a.m. UTC | #3
On Mon, Aug 22, 2016 at 11:35:47AM +0200, Maxime Coquelin wrote:
> >-	virtio_enqueue_offload(m, &virtio_hdr.hdr);
> >-	copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
> >+	/* handle virtio header */
> >+	virtio_hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr;
> >+	virtio_enqueue_offload(mbuf, &(virtio_hdr->hdr));
> Parenthesis around virtio_hdr->hdr shouldn't be needed.
> > 	vhost_log_write(dev, desc->addr, dev->vhost_hlen);
> >-	PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
> Looks like you remove the PRINT_PACKET calls.
> Does it impact performance?

Yes, it does. But it's only enabled for debug mode. Besides that,
it's just a NOOP.

> In any case, it should be mentionned in the commit message.

Agreed. But for this case, we should not remove it: it breaks the
debug-ability.

	--yliu
diff mbox

Patch

diff --git a/lib/librte_vhost/vhost_rxtx.c b/lib/librte_vhost/vhost_rxtx.c
index 08a73fd..b09a9c3 100644
--- a/lib/librte_vhost/vhost_rxtx.c
+++ b/lib/librte_vhost/vhost_rxtx.c
@@ -91,7 +91,7 @@  is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t qp_nb)
 	return (is_tx ^ (idx & 1)) == 0 && idx < qp_nb * VIRTIO_QNUM;
 }
 
-static void
+static inline void __attribute__((always_inline))
 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 {
 	if (m_buf->ol_flags & PKT_TX_L4_MASK) {
@@ -125,427 +125,210 @@  virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 	}
 }
 
-static inline void
-copy_virtio_net_hdr(struct virtio_net *dev, uint64_t desc_addr,
-		    struct virtio_net_hdr_mrg_rxbuf hdr)
+static inline uint32_t __attribute__((always_inline))
+loop_check(struct vhost_virtqueue *vq, uint16_t avail_idx, uint32_t pkt_left)
 {
-	if (dev->vhost_hlen == sizeof(struct virtio_net_hdr_mrg_rxbuf))
-		*(struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr = hdr;
-	else
-		*(struct virtio_net_hdr *)(uintptr_t)desc_addr = hdr.hdr;
+	if (pkt_left == 0 || avail_idx == vq->last_used_idx)
+		return 1;
+
+	return 0;
 }
 
-static inline int __attribute__((always_inline))
-copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
-		  struct rte_mbuf *m, uint16_t desc_idx)
+static inline uint32_t __attribute__((always_inline))
+enqueue_packet(struct virtio_net *dev, struct vhost_virtqueue *vq,
+		uint16_t avail_idx, struct rte_mbuf *mbuf,
+		uint32_t is_mrg_rxbuf)
 {
-	uint32_t desc_avail, desc_offset;
-	uint32_t mbuf_avail, mbuf_offset;
-	uint32_t cpy_len;
+	struct virtio_net_hdr_mrg_rxbuf *virtio_hdr;
 	struct vring_desc *desc;
-	uint64_t desc_addr;
-	struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
-
-	desc = &vq->desc[desc_idx];
+	uint64_t desc_addr = 0;
+	uint32_t desc_chain_head = 0;
+	uint32_t desc_chain_len = 0;
+	uint32_t desc_current = 0;
+	uint32_t desc_offset = 0;
+	uint32_t mbuf_len = 0;
+	uint32_t mbuf_avail = 0;
+	uint32_t copy_len = 0;
+	uint32_t extra_buffers = 0;
+	uint32_t used_idx_round = 0;
+
+	/* start with the first mbuf of the packet */
+	mbuf_len = rte_pktmbuf_data_len(mbuf);
+	mbuf_avail = mbuf_len;
+
+	/* get the current desc */
+	desc_current = vq->avail->ring[(vq->last_used_idx) & (vq->size - 1)];
+	desc_chain_head = desc_current;
+	desc = &vq->desc[desc_current];
 	desc_addr = gpa_to_vva(dev, desc->addr);
-	/*
-	 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
-	 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
-	 * otherwise stores offset on the stack instead of in a register.
-	 */
-	if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr)
-		return -1;
-
-	rte_prefetch0((void *)(uintptr_t)desc_addr);
+	if (unlikely(!desc_addr))
+		goto error;
 
-	virtio_enqueue_offload(m, &virtio_hdr.hdr);
-	copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
+	/* handle virtio header */
+	virtio_hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr;
+	virtio_enqueue_offload(mbuf, &(virtio_hdr->hdr));
 	vhost_log_write(dev, desc->addr, dev->vhost_hlen);
-	PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
-
 	desc_offset = dev->vhost_hlen;
-	desc_avail  = desc->len - dev->vhost_hlen;
+	desc_chain_len = desc_offset;
+	desc_addr += desc_offset;
+	if (is_mrg_rxbuf)
+		virtio_hdr->num_buffers = 1;
 
-	mbuf_avail  = rte_pktmbuf_data_len(m);
-	mbuf_offset = 0;
-	while (mbuf_avail != 0 || m->next != NULL) {
-		/* done with current mbuf, fetch next */
-		if (mbuf_avail == 0) {
-			m = m->next;
-
-			mbuf_offset = 0;
-			mbuf_avail  = rte_pktmbuf_data_len(m);
+	/* start copy from mbuf to desc */
+	while (1) {
+		/* get the next mbuf if the current done */
+		if (!mbuf_avail) {
+			if (mbuf->next) {
+				mbuf = mbuf->next;
+				mbuf_len = rte_pktmbuf_data_len(mbuf);
+				mbuf_avail = mbuf_len;
+			} else
+				break;
 		}
 
-		/* done with current desc buf, fetch next */
-		if (desc_avail == 0) {
-			if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
-				/* Room in vring buffer is not enough */
-				return -1;
-			}
-			if (unlikely(desc->next >= vq->size))
-				return -1;
-
-			desc = &vq->desc[desc->next];
-			desc_addr = gpa_to_vva(dev, desc->addr);
-			if (unlikely(!desc_addr))
-				return -1;
-
-			desc_offset = 0;
-			desc_avail  = desc->len;
+		/* get the next desc if the current done */
+		if (desc->len <= desc_offset) {
+			if (desc->flags & VRING_DESC_F_NEXT) {
+				/* go on with the current desc chain */
+				desc_offset = 0;
+				desc_current = desc->next;
+				desc = &vq->desc[desc_current];
+				desc_addr = gpa_to_vva(dev, desc->addr);
+				if (unlikely(!desc_addr))
+					goto rollback;
+			} else if (is_mrg_rxbuf) {
+				/* start with the next desc chain */
+				used_idx_round = vq->last_used_idx
+					& (vq->size - 1);
+				vq->used->ring[used_idx_round].id =
+					desc_chain_head;
+				vq->used->ring[used_idx_round].len =
+					desc_chain_len;
+				vhost_log_used_vring(dev, vq,
+					offsetof(struct vring_used,
+						ring[used_idx_round]),
+					sizeof(vq->used->ring[
+						used_idx_round]));
+				vq->last_used_idx++;
+				extra_buffers++;
+				virtio_hdr->num_buffers++;
+				if (avail_idx == vq->last_used_idx)
+					goto rollback;
+
+				desc_current =
+					vq->avail->ring[(vq->last_used_idx) &
+					(vq->size - 1)];
+				desc_chain_head = desc_current;
+				desc = &vq->desc[desc_current];
+				desc_addr = gpa_to_vva(dev, desc->addr);
+				if (unlikely(!desc_addr))
+					goto rollback;
+
+				desc_chain_len = 0;
+				desc_offset = 0;
+			} else
+				goto rollback;
 		}
 
-		cpy_len = RTE_MIN(desc_avail, mbuf_avail);
-		rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
-			rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
-			cpy_len);
-		vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
-		PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
-			     cpy_len, 0);
-
-		mbuf_avail  -= cpy_len;
-		mbuf_offset += cpy_len;
-		desc_avail  -= cpy_len;
-		desc_offset += cpy_len;
-	}
-
-	return 0;
-}
-
-/**
- * This function adds buffers to the virtio devices RX virtqueue. Buffers can
- * be received from the physical port or from another virtio device. A packet
- * count is returned to indicate the number of packets that are succesfully
- * added to the RX queue. This function works when the mbuf is scattered, but
- * it doesn't support the mergeable feature.
- */
-static inline uint32_t __attribute__((always_inline))
-virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
-	      struct rte_mbuf **pkts, uint32_t count)
-{
-	struct vhost_virtqueue *vq;
-	uint16_t avail_idx, free_entries, start_idx;
-	uint16_t desc_indexes[MAX_PKT_BURST];
-	uint16_t used_idx;
-	uint32_t i;
-
-	LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
-	if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
-		RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
-			dev->vid, __func__, queue_id);
-		return 0;
+		/* copy mbuf data */
+		copy_len = RTE_MIN(desc->len - desc_offset, mbuf_avail);
+		rte_memcpy((void *)(uintptr_t)desc_addr,
+				rte_pktmbuf_mtod_offset(mbuf, void *,
+					mbuf_len - mbuf_avail),
+				copy_len);
+		vhost_log_write(dev, desc->addr + desc_offset, copy_len);
+		mbuf_avail -= copy_len;
+		desc_offset += copy_len;
+		desc_addr += copy_len;
+		desc_chain_len += copy_len;
 	}
 
-	vq = dev->virtqueue[queue_id];
-	if (unlikely(vq->enabled == 0))
-		return 0;
-
-	avail_idx = *((volatile uint16_t *)&vq->avail->idx);
-	start_idx = vq->last_used_idx;
-	free_entries = avail_idx - start_idx;
-	count = RTE_MIN(count, free_entries);
-	count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
-	if (count == 0)
-		return 0;
-
-	LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
-		dev->vid, start_idx, start_idx + count);
-
-	/* Retrieve all of the desc indexes first to avoid caching issues. */
-	rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
-	for (i = 0; i < count; i++) {
-		used_idx = (start_idx + i) & (vq->size - 1);
-		desc_indexes[i] = vq->avail->ring[used_idx];
-		vq->used->ring[used_idx].id = desc_indexes[i];
-		vq->used->ring[used_idx].len = pkts[i]->pkt_len +
-					       dev->vhost_hlen;
-		vhost_log_used_vring(dev, vq,
-			offsetof(struct vring_used, ring[used_idx]),
-			sizeof(vq->used->ring[used_idx]));
-	}
+	used_idx_round = vq->last_used_idx & (vq->size - 1);
+	vq->used->ring[used_idx_round].id = desc_chain_head;
+	vq->used->ring[used_idx_round].len = desc_chain_len;
+	vhost_log_used_vring(dev, vq,
+		offsetof(struct vring_used, ring[used_idx_round]),
+		sizeof(vq->used->ring[used_idx_round]));
+	vq->last_used_idx++;
 
-	rte_prefetch0(&vq->desc[desc_indexes[0]]);
-	for (i = 0; i < count; i++) {
-		uint16_t desc_idx = desc_indexes[i];
-		int err;
+	return 0;
 
-		err = copy_mbuf_to_desc(dev, vq, pkts[i], desc_idx);
-		if (unlikely(err)) {
-			used_idx = (start_idx + i) & (vq->size - 1);
-			vq->used->ring[used_idx].len = dev->vhost_hlen;
-			vhost_log_used_vring(dev, vq,
-				offsetof(struct vring_used, ring[used_idx]),
-				sizeof(vq->used->ring[used_idx]));
-		}
+rollback:
+	/* rollback on any error if last_used_idx update on-the-fly */
+	if (is_mrg_rxbuf)
+		vq->last_used_idx -= extra_buffers;
 
-		if (i + 1 < count)
-			rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
-	}
+error:
+	return 1;
+}
 
+static inline void __attribute__((always_inline))
+notify_guest(struct virtio_net *dev, struct vhost_virtqueue *vq)
+{
 	rte_smp_wmb();
-
-	*(volatile uint16_t *)&vq->used->idx += count;
-	vq->last_used_idx += count;
-	vhost_log_used_vring(dev, vq,
-		offsetof(struct vring_used, idx),
-		sizeof(vq->used->idx));
-
-	/* flush used->idx update before we read avail->flags. */
+	vq->used->idx = vq->last_used_idx;
+	vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
+			sizeof(vq->used->idx));
 	rte_mb();
-
-	/* Kick the guest if necessary. */
 	if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 			&& (vq->callfd >= 0))
 		eventfd_write(vq->callfd, (eventfd_t)1);
-	return count;
-}
-
-static inline int
-fill_vec_buf(struct vhost_virtqueue *vq, uint32_t avail_idx,
-	     uint32_t *allocated, uint32_t *vec_idx,
-	     struct buf_vector *buf_vec)
-{
-	uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
-	uint32_t vec_id = *vec_idx;
-	uint32_t len    = *allocated;
-
-	while (1) {
-		if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
-			return -1;
-
-		len += vq->desc[idx].len;
-		buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
-		buf_vec[vec_id].buf_len  = vq->desc[idx].len;
-		buf_vec[vec_id].desc_idx = idx;
-		vec_id++;
-
-		if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
-			break;
-
-		idx = vq->desc[idx].next;
-	}
-
-	*allocated = len;
-	*vec_idx   = vec_id;
-
-	return 0;
-}
-
-/*
- * Returns -1 on fail, 0 on success
- */
-static inline int
-reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
-			    uint16_t *end, struct buf_vector *buf_vec)
-{
-	uint16_t cur_idx;
-	uint16_t avail_idx;
-	uint32_t allocated = 0;
-	uint32_t vec_idx = 0;
-	uint16_t tries = 0;
-
-	cur_idx  = vq->last_used_idx;
-
-	while (1) {
-		avail_idx = *((volatile uint16_t *)&vq->avail->idx);
-		if (unlikely(cur_idx == avail_idx))
-			return -1;
-
-		if (unlikely(fill_vec_buf(vq, cur_idx, &allocated,
-					  &vec_idx, buf_vec) < 0))
-			return -1;
-
-		cur_idx++;
-		tries++;
-
-		if (allocated >= size)
-			break;
-
-		/*
-		 * if we tried all available ring items, and still
-		 * can't get enough buf, it means something abnormal
-		 * happened.
-		 */
-		if (unlikely(tries >= vq->size))
-			return -1;
-	}
-
-	*end = cur_idx;
-	return 0;
 }
 
-static inline uint32_t __attribute__((always_inline))
-copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
-			    uint16_t end_idx, struct rte_mbuf *m,
-			    struct buf_vector *buf_vec)
+uint16_t
+rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
+	struct rte_mbuf **pkts, uint16_t count)
 {
-	struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
-	uint32_t vec_idx = 0;
-	uint16_t start_idx = vq->last_used_idx;
-	uint16_t cur_idx = start_idx;
-	uint64_t desc_addr;
-	uint32_t mbuf_offset, mbuf_avail;
-	uint32_t desc_offset, desc_avail;
-	uint32_t cpy_len;
-	uint16_t desc_idx, used_idx;
-
-	if (unlikely(m == NULL))
+	struct vhost_virtqueue *vq;
+	struct virtio_net *dev;
+	uint32_t pkt_idx = 0;
+	uint32_t pkt_left = 0;
+	uint32_t pkt_sent = 0;
+	uint32_t is_mrg_rxbuf = 0;
+	uint16_t avail_idx = 0;
+
+	/* precheck */
+	if (unlikely(count == 0))
 		return 0;
 
-	LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
-		dev->vid, cur_idx, end_idx);
+	count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 
-	desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
-	if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr)
+	dev = get_device(vid);
+	if (unlikely(!dev))
 		return 0;
 
-	rte_prefetch0((void *)(uintptr_t)desc_addr);
-
-	virtio_hdr.num_buffers = end_idx - start_idx;
-	LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
-		dev->vid, virtio_hdr.num_buffers);
-
-	virtio_enqueue_offload(m, &virtio_hdr.hdr);
-	copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
-	vhost_log_write(dev, buf_vec[vec_idx].buf_addr, dev->vhost_hlen);
-	PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
-
-	desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
-	desc_offset = dev->vhost_hlen;
-
-	mbuf_avail  = rte_pktmbuf_data_len(m);
-	mbuf_offset = 0;
-	while (mbuf_avail != 0 || m->next != NULL) {
-		/* done with current desc buf, get the next one */
-		if (desc_avail == 0) {
-			desc_idx = buf_vec[vec_idx].desc_idx;
-
-			if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
-				/* Update used ring with desc information */
-				used_idx = cur_idx++ & (vq->size - 1);
-				vq->used->ring[used_idx].id  = desc_idx;
-				vq->used->ring[used_idx].len = desc_offset;
-				vhost_log_used_vring(dev, vq,
-					offsetof(struct vring_used,
-						 ring[used_idx]),
-					sizeof(vq->used->ring[used_idx]));
-			}
-
-			vec_idx++;
-			desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
-			if (unlikely(!desc_addr))
-				return 0;
-
-			/* Prefetch buffer address. */
-			rte_prefetch0((void *)(uintptr_t)desc_addr);
-			desc_offset = 0;
-			desc_avail  = buf_vec[vec_idx].buf_len;
-		}
-
-		/* done with current mbuf, get the next one */
-		if (mbuf_avail == 0) {
-			m = m->next;
-
-			mbuf_offset = 0;
-			mbuf_avail  = rte_pktmbuf_data_len(m);
-		}
-
-		cpy_len = RTE_MIN(desc_avail, mbuf_avail);
-		rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
-			rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
-			cpy_len);
-		vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
-			cpy_len);
-		PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
-			cpy_len, 0);
-
-		mbuf_avail  -= cpy_len;
-		mbuf_offset += cpy_len;
-		desc_avail  -= cpy_len;
-		desc_offset += cpy_len;
-	}
-
-	used_idx = cur_idx & (vq->size - 1);
-	vq->used->ring[used_idx].id = buf_vec[vec_idx].desc_idx;
-	vq->used->ring[used_idx].len = desc_offset;
-	vhost_log_used_vring(dev, vq,
-		offsetof(struct vring_used, ring[used_idx]),
-		sizeof(vq->used->ring[used_idx]));
-
-	return end_idx - start_idx;
-}
-
-static inline uint32_t __attribute__((always_inline))
-virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
-	struct rte_mbuf **pkts, uint32_t count)
-{
-	struct vhost_virtqueue *vq;
-	uint32_t pkt_idx = 0, nr_used = 0;
-	uint16_t end;
-	struct buf_vector buf_vec[BUF_VECTOR_MAX];
-
-	LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
-	if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
-		RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
-			dev->vid, __func__, queue_id);
+	if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb)))
 		return 0;
-	}
 
 	vq = dev->virtqueue[queue_id];
-	if (unlikely(vq->enabled == 0))
+	if (unlikely(!vq->enabled))
 		return 0;
 
-	count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
-	if (count == 0)
-		return 0;
+	if (dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF))
+		is_mrg_rxbuf = 1;
 
-	for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
-		uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
-
-		if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len,
-							 &end, buf_vec) < 0)) {
-			LOG_DEBUG(VHOST_DATA,
-				"(%d) failed to get enough desc from vring\n",
-				dev->vid);
+	/* start enqueuing packets 1 by 1 */
+	pkt_idx = 0;
+	pkt_left = count;
+	avail_idx = *((volatile uint16_t *)&vq->avail->idx);
+	while (1) {
+		if (loop_check(vq, avail_idx, pkt_left))
 			break;
-		}
-
-		nr_used = copy_mbuf_to_desc_mergeable(dev, vq, end,
-						      pkts[pkt_idx], buf_vec);
-		rte_smp_wmb();
 
-		*(volatile uint16_t *)&vq->used->idx += nr_used;
-		vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
-			sizeof(vq->used->idx));
-		vq->last_used_idx += nr_used;
-	}
-
-	if (likely(pkt_idx)) {
-		/* flush used->idx update before we read avail->flags. */
-		rte_mb();
+		if (enqueue_packet(dev, vq, avail_idx, pkts[pkt_idx],
+					is_mrg_rxbuf))
+			break;
 
-		/* Kick the guest if necessary. */
-		if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
-				&& (vq->callfd >= 0))
-			eventfd_write(vq->callfd, (eventfd_t)1);
+		pkt_idx++;
+		pkt_sent++;
+		pkt_left--;
 	}
 
-	return pkt_idx;
-}
-
-uint16_t
-rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
-	struct rte_mbuf **pkts, uint16_t count)
-{
-	struct virtio_net *dev = get_device(vid);
-
-	if (!dev)
-		return 0;
+	/* update used idx and kick the guest if necessary */
+	if (pkt_sent)
+		notify_guest(dev, vq);
 
-	if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
-		return virtio_dev_merge_rx(dev, queue_id, pkts, count);
-	else
-		return virtio_dev_rx(dev, queue_id, pkts, count);
+	return pkt_sent;
 }
 
 static void