Re: [dpdk-dev] [PATCH 5/7] net/virtio: support IN_ORDER Rx and Tx

[Tiwei Bie <tiwei.bie@intel.com>]

On Fri, Jun 08, 2018 at 05:07:22PM +0800, Marvin Liu wrote:
> IN_ORDER Rx function can support merge-able feature. Descriptors
> allocation and free will be done in bulk.
> 
> Virtio dequeue logic:
>     dequeue_burst_rx(burst mbufs)
>     for (each mbuf b) {
>             if (b need merge) {
>                     merge remained mbufs
>                     add merged mbuf to return mbufs list
>             } else {
>                     add mbuf to return mbufs list
>             }
>     }
>     if (last mbuf c need merge) {
>             dequeue_burst_rx(required mbufs)
>             merge last mbuf c
>     }
>     refill_avail_ring_bulk()
>     update_avail_ring()
>     return mbufs list
> 
> IN_ORDER Tx function can support indirect and offloading features.
> 
> Virtio enqueue logic:
>     xmit_cleanup(used descs)
>     for (each xmit mbuf b) {
>             if (b can inorder xmit) {
>                     add mbuf b to inorder burst list
>             } else {
>                     xmit inorder burst list
>                     xmit mbuf b with normal xmit
>             }
>     }
>     if (inorder burst list not empty) {
>             xmit inorder burst list
>     }
>     update_avail_ring()
> 
> Signed-off-by: Marvin Liu <yong.liu@intel.com>
> 
> diff --git a/drivers/net/virtio/virtio_ethdev.h b/drivers/net/virtio/virtio_ethdev.h
> index bb40064ea..25697c872 100644
> --- a/drivers/net/virtio/virtio_ethdev.h
> +++ b/drivers/net/virtio/virtio_ethdev.h
> @@ -83,9 +83,15 @@ uint16_t virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
>  uint16_t virtio_recv_mergeable_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
>  		uint16_t nb_pkts);
>  
> +uint16_t virtio_recv_inorder_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
> +		uint16_t nb_pkts);
> +
>  uint16_t virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
>  		uint16_t nb_pkts);
>  
> +uint16_t virtio_xmit_inorder_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
> +		uint16_t nb_pkts);
> +
>  uint16_t virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
>  		uint16_t nb_pkts);
>  
> diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
> index 0bca29855..d0473d6b4 100644
> --- a/drivers/net/virtio/virtio_rxtx.c
> +++ b/drivers/net/virtio/virtio_rxtx.c
> @@ -122,6 +122,45 @@ virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
>  	return i;
>  }
>  
> +static uint16_t
> +virtqueue_dequeue_inorder_rx(struct virtqueue *vq,

virtqueue_dequeue_burst_rx_inorder() will be better.

> +			struct rte_mbuf **rx_pkts,
> +			uint32_t *len,
> +			uint16_t num)
> +{
> +	struct vring_used_elem *uep;
> +	struct rte_mbuf *cookie;
> +	uint16_t used_idx;
> +	uint16_t i;
> +
> +	if (unlikely(num == 0))
> +		return 0;
> +
> +	used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);

Above assignment is misleading.

> +	for (i = 0; i < num; i++) {
> +		used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
> +		/* Desc idx same as used idx */
> +		uep = &vq->vq_ring.used->ring[used_idx];
> +		len[i] = uep->len;
> +		cookie = (struct rte_mbuf *)vq->vq_descx[used_idx].cookie;
> +
> +		if (unlikely(cookie == NULL)) {
> +			PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
> +				vq->vq_used_cons_idx);
> +			break;
> +		}
> +
> +		rte_prefetch0(cookie);
> +		rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
> +		rx_pkts[i]  = cookie;
> +		vq->vq_used_cons_idx++;
> +		vq->vq_descx[used_idx].cookie = NULL;
> +	}
> +
> +	vq_ring_free_inorder(vq, used_idx, i);
> +	return i;
> +}
> +
>  #ifndef DEFAULT_TX_FREE_THRESH
>  #define DEFAULT_TX_FREE_THRESH 32
>  #endif
> @@ -150,6 +189,70 @@ virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
>  	}
>  }
>  
> +/* Cleanup from completed inorder transmits. */
> +static void
> +virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num)
> +{
> +	uint16_t i, used_idx;
> +
> +	used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);

Above assignment is misleading.

> +	for (i = 0; i < num; i++) {
> +		struct vq_desc_extra *dxp;
> +
> +		used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
> +		dxp = &vq->vq_descx[used_idx];
> +		vq->vq_used_cons_idx++;
> +
> +		if (dxp->cookie != NULL) {
> +			rte_pktmbuf_free(dxp->cookie);
> +			dxp->cookie = NULL;
> +		}
> +	}
> +	vq_ring_free_inorder(vq, used_idx, num);

If num is zero, we can't free used_idx.

> +}
> +
> +static inline int
> +virtqueue_enqueue_inorder_refill(struct virtqueue *vq,

virtqueue_enqueue_recv_refill_inorder() will be better.


> +			struct rte_mbuf **cookies,
> +			uint16_t num)
> +{
> +	struct vq_desc_extra *dxp;
> +	struct virtio_hw *hw = vq->hw;
> +	struct vring_desc *start_dp;
> +	uint16_t head_idx, idx, i = 0;
> +
> +	if (unlikely(vq->vq_free_cnt == 0))
> +		return -ENOSPC;
> +	if (unlikely(vq->vq_free_cnt < num))
> +		return -EMSGSIZE;
> +
> +	head_idx = vq->vq_desc_head_idx & (vq->vq_nentries - 1);
> +	start_dp = vq->vq_ring.desc;
> +
> +	while (i < num) {
> +		idx = head_idx & (vq->vq_nentries - 1);
> +		dxp = &vq->vq_descx[idx];
> +		dxp->cookie = (void *)(struct rte_mbuf *)cookies[i];

Casting isn't necessary.

> +		dxp->ndescs = 1;
> +
> +		start_dp[idx].addr =
> +				VIRTIO_MBUF_ADDR(cookies[i], vq) +
> +				RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
> +		start_dp[idx].len =
> +				cookies[i]->buf_len -
> +				RTE_PKTMBUF_HEADROOM +
> +				hw->vtnet_hdr_size;
> +		start_dp[idx].flags =  VRING_DESC_F_WRITE;
> +
> +		head_idx++;
> +		i++;
> +	}
> +
> +	vq->vq_avail_idx += num;
> +	vq->vq_desc_head_idx += num;
> +	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
> +	return 0;
> +}
>  
>  static inline int
>  virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
> @@ -246,9 +349,113 @@ tx_offload_enabled(struct virtio_hw *hw)
>  		(var) = (val);			\
>  } while (0)
>  
[...]
> +
> +static void
> +virtio_discard_inorder_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)

virtio_discard_rxbuf_inorder() will be better.

> +{
> +	int error;
> +
> +	error = virtqueue_enqueue_inorder_refill(vq, &m, 1);
>  	if (unlikely(error)) {
>  		RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
>  		rte_pktmbuf_free(m);
> @@ -813,6 +1002,191 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
>  	return nb_rx;
>  }
>  
> +uint16_t
> +virtio_recv_inorder_pkts(void *rx_queue,

virtio_recv_mergeable_pkts_inorder() will be better.

> +			struct rte_mbuf **rx_pkts,
> +			uint16_t nb_pkts)
> +{
> +	struct virtnet_rx *rxvq = rx_queue;
> +	struct virtqueue *vq = rxvq->vq;
> +	struct virtio_hw *hw = vq->hw;
> +	struct rte_mbuf *rxm;
> +	struct rte_mbuf *prev;
> +	uint16_t nb_used, num, nb_rx;
> +	uint32_t len[VIRTIO_MBUF_BURST_SZ];
> +	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
> +	int error;
> +	uint32_t i, nb_enqueued;
> +	uint32_t seg_num;
> +	uint32_t seg_res;
> +	uint32_t hdr_size;
> +	int offload;
> +
> +	nb_rx = 0;
> +	if (unlikely(hw->started == 0))
> +		return nb_rx;
> +
> +	nb_used = VIRTQUEUE_NUSED(vq);
> +	nb_used = RTE_MIN(nb_used, nb_pkts);
> +	nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
> +
> +	virtio_rmb();
> +
> +	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
> +
> +	i = 0;
> +	nb_enqueued = 0;
> +	seg_num = 1;
> +	seg_res = 0;
> +	hdr_size = hw->vtnet_hdr_size;
> +	offload = rx_offload_enabled(hw);
> +
> +	num = virtqueue_dequeue_inorder_rx(vq, rcv_pkts, len, nb_used);
> +
> +	for (; i < num; i++) {

Why not initialize i in for ()?

> +		struct virtio_net_hdr_mrg_rxbuf *header;
> +
> +		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
> +		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
> +
> +		rxm = rcv_pkts[i];
> +
> +		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
> +			PMD_RX_LOG(ERR, "Packet drop");
> +			nb_enqueued++;
> +			virtio_discard_inorder_rxbuf(vq, rxm);
> +			rxvq->stats.errors++;
> +			continue;
> +		}
> +
> +		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
> +			RTE_PKTMBUF_HEADROOM - hdr_size);
> +		seg_num = header->num_buffers;
> +
> +		if (seg_num == 0)
> +			seg_num = 1;
> +
> +		rxm->data_off = RTE_PKTMBUF_HEADROOM;
> +		rxm->nb_segs = seg_num;
> +		rxm->ol_flags = 0;
> +		rxm->vlan_tci = 0;
> +		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
> +		rxm->data_len = (uint16_t)(len[i] - hdr_size);
> +
> +		rxm->port = rxvq->port_id;
> +
> +		rx_pkts[nb_rx] = rxm;
> +		prev = rxm;
> +
> +		if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
> +			virtio_discard_inorder_rxbuf(vq, rxm);
> +			rxvq->stats.errors++;
> +			continue;
> +		}
> +
> +		seg_res = seg_num - 1;
> +
> +		/* Merge remaining segments */
> +		while (seg_res != 0 && i < num) {
> +			rxm = rcv_pkts[i];
> +			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
> +			rxm->pkt_len = (uint32_t)(len[i]);
> +			rxm->data_len = (uint16_t)(len[i]);
> +
> +			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
> +			rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
> +
> +			i++;
> +
> +			if (prev)
> +				prev->next = rxm;
> +
> +			prev = rxm;
> +			seg_res -= 1;
> +		}
> +
> +		if (hw->vlan_strip)
> +			rte_vlan_strip(rx_pkts[nb_rx]);
> +
> +		VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
> +			rx_pkts[nb_rx]->data_len);
> +
> +		rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
> +		virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);

When seg_res != 0, we cannot run above code.

> +		nb_rx++;
> +	}
> +
> +	/* Last packet still need merge segments */
> +	while (seg_res != 0) {
> +		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
> +					VIRTIO_MBUF_BURST_SZ);
> +
> +		prev = rcv_pkts[nb_rx - 1];
> +		if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
> +			num = virtqueue_dequeue_inorder_rx(vq, rcv_pkts, len,
> +							   rcv_cnt);
> +			uint16_t extra_idx = 0;
> +
> +			rcv_cnt = num;
> +			while (extra_idx < rcv_cnt) {
> +				rxm = rcv_pkts[extra_idx];
> +				rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
> +				rxm->pkt_len = (uint32_t)(len[extra_idx]);
> +				rxm->data_len = (uint16_t)(len[extra_idx]);
> +				prev->next = rxm;
> +				prev = rxm;
> +				rx_pkts[nb_rx - 1]->pkt_len += len[extra_idx];
> +				rx_pkts[nb_rx - 1]->data_len += len[extra_idx];
> +				extra_idx += 1;
> +			};
> +			seg_res -= rcv_cnt;
> +		} else {
> +			PMD_RX_LOG(ERR,
> +					"No enough segments for packet.");
> +			virtio_discard_inorder_rxbuf(vq, prev);
> +			rxvq->stats.errors++;
> +			nb_rx--;
> +			break;
> +		}
> +	}
> +
> +	rxvq->stats.packets += nb_rx;
> +
> +	/* Allocate new mbuf for the used descriptor */
> +	error = ENOSPC;

Above assignment is meaningless.

> +
> +	if (likely(!virtqueue_full(vq))) {
> +		/* free_cnt may include mrg descs */
> +		uint16_t free_cnt = vq->vq_free_cnt;
> +		struct rte_mbuf *new_pkts[free_cnt];
> +
> +		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
> +			error = virtqueue_enqueue_inorder_refill(vq, new_pkts,
> +					free_cnt);
> +			if (unlikely(error)) {
> +				for (i = 0; i < free_cnt; i++)
> +					rte_pktmbuf_free(new_pkts[i]);
> +			}
> +			nb_enqueued += free_cnt;
> +		} else {
> +			struct rte_eth_dev *dev
> +				= &rte_eth_devices[rxvq->port_id];
> +			dev->data->rx_mbuf_alloc_failed += free_cnt;
> +		}
> +	}
> +
> +	if (likely(nb_enqueued)) {
> +		vq_update_avail_idx(vq);
> +
> +		if (unlikely(virtqueue_kick_prepare(vq))) {
> +			virtqueue_notify(vq);
> +			PMD_RX_LOG(DEBUG, "Notified");
> +		}
> +	}
> +
> +	return nb_rx;
> +}
> +
>  uint16_t
>  virtio_recv_mergeable_pkts(void *rx_queue,
>  			struct rte_mbuf **rx_pkts,
> @@ -1060,7 +1434,8 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
>  		}
>  
>  		/* Enqueue Packet buffers */
> -		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
> +		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
> +			can_push, 0);
>  
>  		txvq->stats.bytes += txm->pkt_len;
>  		virtio_update_packet_stats(&txvq->stats, txm);
> @@ -1079,3 +1454,121 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
>  
>  	return nb_tx;
>  }
> +
> +uint16_t
> +virtio_xmit_inorder_pkts(void *tx_queue,

virtio_xmit_pkts_inorder() will be better.

> +			struct rte_mbuf **tx_pkts,
> +			uint16_t nb_pkts)
> +{
> +	struct virtnet_tx *txvq = tx_queue;
> +	struct virtqueue *vq = txvq->vq;
> +	struct virtio_hw *hw = vq->hw;
> +	uint16_t hdr_size = hw->vtnet_hdr_size;
> +	uint16_t nb_used, nb_avail, nb_tx = 0, inorder_tx = 0;
> +	struct rte_mbuf *inorder_pkts[nb_pkts];
> +	int error;
> +
> +	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
> +		return nb_tx;
> +
> +	if (unlikely(nb_pkts < 1))
> +		return nb_pkts;
> +
> +	VIRTQUEUE_DUMP(vq);
> +	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
> +	nb_used = VIRTQUEUE_NUSED(vq);
> +
> +	virtio_rmb();
> +	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
> +		virtio_xmit_cleanup_inorder(vq, nb_used);
> +
> +	nb_avail = RTE_MIN(vq->vq_free_cnt, nb_pkts);
> +
> +	for (nb_tx = 0; nb_tx < nb_avail; nb_tx++) {
> +		struct rte_mbuf *txm = tx_pkts[nb_tx];
> +		int slots, need;
> +		int can_push = 0, use_indirect = 0, flush_inorder = 1;
> +
> +		/* Do VLAN tag insertion */
> +		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
> +			error = rte_vlan_insert(&txm);
> +			if (unlikely(error)) {
> +				rte_pktmbuf_free(txm);
> +				continue;
> +			}
> +		}
> +
> +		txvq->stats.bytes += txm->pkt_len;
> +		virtio_update_packet_stats(&txvq->stats, txm);
> +
> +		/* optimize ring usage */
> +		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
> +		     vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
> +		     rte_mbuf_refcnt_read(txm) == 1 &&
> +		     RTE_MBUF_DIRECT(txm) &&
> +		     txm->nb_segs == 1 &&
> +		     rte_pktmbuf_headroom(txm) >= hdr_size &&
> +		     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
> +				__alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
> +			can_push = 1;

There is no need to have 'can_push' any more.

> +
> +			inorder_pkts[inorder_tx] = txm;
> +			inorder_tx++;

inorder_tx isn't a good name.

> +			flush_inorder = 0;

I think there is no need to introduce 'flush_inorder'.
And it will make the code much more readable by writing
the code like:

	for (nb_tx = 0; nb_tx < nb_avail; nb_tx++) {
		struct rte_mbuf *txm = tx_pkts[nb_tx];

		......

		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
		     vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
		     rte_mbuf_refcnt_read(txm) == 1 &&
		     RTE_MBUF_DIRECT(txm) &&
		     txm->nb_segs == 1 &&
		     rte_pktmbuf_headroom(txm) >= hdr_size &&
		     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
				__alignof__(struct virtio_net_hdr_mrg_rxbuf))) {

			inorder_pkts[nb_inorder_pkts++] = txm;
			continue;
		}

		/* Flush inorder packets if any. */
		if (nb_inorder_pkts) {
			virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
				nb_inorder_pkts);
			nb_inorder_pkts = 0;
		}

		......

		virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
	}

	/* Flush inorder packets if any. */
	if (nb_inorder_pkts) {
		virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
			nb_inorder_pkts);
		nb_inorder_pkts = 0;
	}



> +		} else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
> +				txm->nb_segs < VIRTIO_MAX_TX_INDIRECT) {
> +			use_indirect = 1;
> +		}

There is no need to try to use INDIRECT.

> +
> +		if (flush_inorder) {
> +			/* Flush inorder packets */
> +			virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
> +				inorder_tx);
> +			inorder_tx = 0;
> +
> +			/* How many main ring entries are needed to this Tx?
> +			 * any_layout => number of segments
> +			 * indirect   => 1
> +			 * default    => number of segments + 1
> +			 */
> +			slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
> +			need = slots - vq->vq_free_cnt;
> +
> +			if (unlikely(need > 0)) {
> +				nb_used = VIRTQUEUE_NUSED(vq);
> +				virtio_rmb();
> +				need = RTE_MIN(need, (int)nb_used);
> +
> +				virtio_xmit_cleanup_inorder(vq, need);
> +				need = slots - vq->vq_free_cnt;
> +				if (unlikely(need > 0)) {
> +					PMD_TX_LOG(ERR,
> +						"No free tx descriptors to transmit");
> +					break;
> +				}
> +			}
> +			/* Enqueue Packet buffers */
> +			virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
> +				can_push, 1);
> +		}
> +	}
> +
> +	/* Transmit all inorder packets */
> +	if (inorder_tx)
> +		virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts, inorder_tx);
> +
> +	txvq->stats.packets += nb_tx;
> +
> +	if (likely(nb_tx)) {
> +		vq_update_avail_idx(vq);
> +
> +		if (unlikely(virtqueue_kick_prepare(vq))) {
> +			virtqueue_notify(vq);
> +			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
> +		}
> +	}
> +
> +	VIRTQUEUE_DUMP(vq);
> +
> +	return nb_tx;
> +}
> -- 
> 2.17.0
>