examples/vhost: sample code refactor

index 71d8a1f1f5..b0b04aa453 100644
@@ -17,6 +17,7 @@ struct packet_tracker {
 	unsigned short next_read;
 	unsigned short next_write;
 	unsigned short last_remain;
+	unsigned short ioat_space;
 };
 
 struct packet_tracker cb_tracker[MAX_VHOST_DEVICE];

@@ -113,7 +114,7 @@ open_ioat(const char *value)
 			goto out;
 		}
 		rte_rawdev_start(dev_id);
-
+		cb_tracker[dev_id].ioat_space = IOAT_RING_SIZE;
 		dma_info->nr++;
 		i++;
 	}

@@ -140,13 +141,9 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 			src = descs[i_desc].src;
 			dst = descs[i_desc].dst;
 			i_seg = 0;
+			if (cb_tracker[dev_id].ioat_space < src->nr_segs)
+				break;
 			while (i_seg < src->nr_segs) {
-				/*
-				 * TODO: Assuming that the ring space of the
-				 * IOAT device is large enough, so there is no
-				 * error here, and the actual error handling
-				 * will be added later.
-				 */
 				rte_ioat_enqueue_copy(dev_id,
 					(uintptr_t)(src->iov[i_seg].iov_base)
 						+ src->offset,

@@ -158,7 +155,8 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 				i_seg++;
 			}
 			write &= mask;
-			cb_tracker[dev_id].size_track[write] = i_seg;
+			cb_tracker[dev_id].size_track[write] = src->nr_segs;
+			cb_tracker[dev_id].ioat_space -= src->nr_segs;
 			write++;
 		}
 	} else {

@@ -186,6 +184,7 @@ ioat_check_completed_copies_cb(int vid, uint16_t queue_id,
 		int dev_id = dma_bind[vid].dmas[queue_id * 2
 				+ VIRTIO_RXQ].dev_id;
 		n_seg = rte_ioat_completed_ops(dev_id, 255, dump, dump);
+		cb_tracker[dev_id].ioat_space += n_seg;
 		n_seg += cb_tracker[dev_id].last_remain;
 		if (!n_seg)
 			return 0;

index 8d8c3038bf..28226a4ff7 100644
@@ -182,6 +182,11 @@ struct mbuf_table {
 /* TX queue for each data core. */
 struct mbuf_table lcore_tx_queue[RTE_MAX_LCORE];
 
+static uint64_t vhost_tsc[MAX_VHOST_DEVICE];
+
+/* TX queue for each vhost device. */
+struct mbuf_table vhost_m_table[MAX_VHOST_DEVICE];
+
 #define MBUF_TABLE_DRAIN_TSC	((rte_get_tsc_hz() + US_PER_S - 1) \
 				 / US_PER_S * BURST_TX_DRAIN_US)
 #define VLAN_HLEN       4

@@ -804,6 +809,13 @@ unlink_vmdq(struct vhost_dev *vdev)
 	}
 }
 
+static inline void
+free_pkts(struct rte_mbuf **pkts, uint16_t n)
+{
+	while (n--)
+		rte_pktmbuf_free(pkts[n]);
+}
+
 static __rte_always_inline void
 virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
 	    struct rte_mbuf *m)

@@ -837,6 +849,40 @@ virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
 	}
 }
 
+static __rte_always_inline void
+drain_vhost(struct vhost_dev *dst_vdev, struct rte_mbuf **m, uint16_t nr_xmit)
+{
+	uint16_t ret, nr_cpl;
+	struct rte_mbuf *m_cpl[MAX_PKT_BURST];
+
+	if (builtin_net_driver) {
+		ret = vs_enqueue_pkts(dst_vdev, VIRTIO_RXQ, m, nr_xmit);
+	} else if (async_vhost_driver) {
+		ret = rte_vhost_submit_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ,
+						m, nr_xmit);
+		dst_vdev->nr_async_pkts += ret;
+		free_pkts(&m[ret], nr_xmit - ret);
+
+		while (likely(dst_vdev->nr_async_pkts)) {
+			nr_cpl = rte_vhost_poll_enqueue_completed(dst_vdev->vid,
+					VIRTIO_RXQ, m_cpl, MAX_PKT_BURST);
+			dst_vdev->nr_async_pkts -= nr_cpl;
+			free_pkts(m_cpl, nr_cpl);
+		}
+	} else {
+		ret = rte_vhost_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ,
+						m, nr_xmit);
+	}
+
+	if (enable_stats) {
+		rte_atomic64_add(&dst_vdev->stats.rx_total_atomic, nr_xmit);
+		rte_atomic64_add(&dst_vdev->stats.rx_atomic, ret);
+	}
+
+	if (!async_vhost_driver)
+		free_pkts(m, nr_xmit);
+}
+
 /*
  * Check if the packet destination MAC address is for a local device. If so then put
  * the packet on that devices RX queue. If not then return.

@@ -846,6 +892,7 @@ virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 {
 	struct rte_ether_hdr *pkt_hdr;
 	struct vhost_dev *dst_vdev;
+	struct mbuf_table *vhost_txq;
 
 	pkt_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
 

@@ -869,7 +916,19 @@ virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 		return 0;
 	}
 
-	virtio_xmit(dst_vdev, vdev, m);
+	vhost_txq = &vhost_m_table[dst_vdev->vid];
+	vhost_txq->m_table[vhost_txq->len++] = m;
+
+	if (enable_stats) {
+		vdev->stats.tx_total++;
+		vdev->stats.tx++;
+	}
+
+	if (unlikely(vhost_txq->len == MAX_PKT_BURST)) {
+		drain_vhost(dst_vdev, vhost_txq->m_table, MAX_PKT_BURST);
+		vhost_txq->len = 0;
+		vhost_tsc[dst_vdev->vid] = rte_rdtsc();
+	}
 	return 0;
 }
 

@@ -940,13 +999,6 @@ static void virtio_tx_offload(struct rte_mbuf *m)
 	tcp_hdr->cksum = get_psd_sum(l3_hdr, m->ol_flags);
 }
 
-static inline void
-free_pkts(struct rte_mbuf **pkts, uint16_t n)
-{
-	while (n--)
-		rte_pktmbuf_free(pkts[n]);
-}
-
 static __rte_always_inline void
 do_drain_mbuf_table(struct mbuf_table *tx_q)
 {

@@ -986,7 +1038,6 @@ virtio_tx_route(struct vhost_dev *vdev, struct rte_mbuf *m, uint16_t vlan_tag)
 
 	/*check if destination is local VM*/
 	if ((vm2vm_mode == VM2VM_SOFTWARE) && (virtio_tx_local(vdev, m) == 0)) {
-		rte_pktmbuf_free(m);
 		return;
 	}
 

@@ -1144,8 +1195,10 @@ static __rte_always_inline void
 drain_virtio_tx(struct vhost_dev *vdev)
 {
 	struct rte_mbuf *pkts[MAX_PKT_BURST];
+	struct mbuf_table *vhost_txq;
 	uint16_t count;
 	uint16_t i;
+	uint64_t cur_tsc;
 
 	if (builtin_net_driver) {
 		count = vs_dequeue_pkts(vdev, VIRTIO_TXQ, mbuf_pool,

@@ -1163,6 +1216,19 @@ drain_virtio_tx(struct vhost_dev *vdev)
 
 	for (i = 0; i < count; ++i)
 		virtio_tx_route(vdev, pkts[i], vlan_tags[vdev->vid]);
+
+	vhost_txq = &vhost_m_table[vdev->vid];
+	cur_tsc = rte_rdtsc();
+	if (unlikely(cur_tsc - vhost_tsc[vdev->vid] > MBUF_TABLE_DRAIN_TSC)) {
+		if (vhost_txq->len) {
+			RTE_LOG_DP(DEBUG, VHOST_DATA,
+				"Vhost tX queue drained after timeout with burst size %u\n",
+				vhost_txq->len);
+			drain_vhost(vdev, vhost_txq->m_table, vhost_txq->len);
+			vhost_txq->len = 0;
+			vhost_tsc[vdev->vid] = cur_tsc;
+		}
+	}
 }
 
 /*

index 28226a4ff7..0113147876 100644
@@ -817,26 +817,26 @@ free_pkts(struct rte_mbuf **pkts, uint16_t n)
 }
 
 static __rte_always_inline void
-virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
+complete_async_pkts(struct vhost_dev *vdev)
+{
+	struct rte_mbuf *p_cpl[MAX_PKT_BURST];
+	uint16_t complete_count;
+
+	complete_count = rte_vhost_poll_enqueue_completed(vdev->vid,
+					VIRTIO_RXQ, p_cpl, MAX_PKT_BURST);
+	rte_atomic16_sub(&vdev->nr_async_pkts, complete_count);
+	if (complete_count)
+		free_pkts(p_cpl, complete_count);
+}
+
+static __rte_always_inline void
+sync_virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
 	    struct rte_mbuf *m)
 {
 	uint16_t ret;
-	struct rte_mbuf *m_cpl[1];
 
 	if (builtin_net_driver) {
 		ret = vs_enqueue_pkts(dst_vdev, VIRTIO_RXQ, &m, 1);
-	} else if (async_vhost_driver) {
-		ret = rte_vhost_submit_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ,
-						&m, 1);
-
-		if (likely(ret))
-			dst_vdev->nr_async_pkts++;
-
-		while (likely(dst_vdev->nr_async_pkts)) {
-			if (rte_vhost_poll_enqueue_completed(dst_vdev->vid,
-					VIRTIO_RXQ, m_cpl, 1))
-				dst_vdev->nr_async_pkts--;
-		}
 	} else {
 		ret = rte_vhost_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ, &m, 1);
 	}

@@ -850,25 +850,25 @@ virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
 }
 
 static __rte_always_inline void
-drain_vhost(struct vhost_dev *dst_vdev, struct rte_mbuf **m, uint16_t nr_xmit)
+drain_vhost(struct vhost_dev *dst_vdev)
 {
-	uint16_t ret, nr_cpl;
-	struct rte_mbuf *m_cpl[MAX_PKT_BURST];
+	uint16_t ret;
+	uint16_t nr_xmit = vhost_m_table[dst_vdev->vid].len;
+	struct rte_mbuf **m = vhost_m_table[dst_vdev->vid].m_table;
 
 	if (builtin_net_driver) {
 		ret = vs_enqueue_pkts(dst_vdev, VIRTIO_RXQ, m, nr_xmit);
 	} else if (async_vhost_driver) {
+		uint32_t cpu_cpl_nr;
+		struct rte_mbuf *m_cpu_cpl[nr_xmit];
+		complete_async_pkts(dst_vdev);
+		while (rte_atomic16_read(&dst_vdev->nr_async_pkts) >= 128)
+			complete_async_pkts(dst_vdev);
+
 		ret = rte_vhost_submit_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ,
-						m, nr_xmit);
-		dst_vdev->nr_async_pkts += ret;
+					m, nr_xmit, m_cpu_cpl, &cpu_cpl_nr);
+		rte_atomic16_add(&dst_vdev->nr_async_pkts, ret - cpu_cpl_nr);
 		free_pkts(&m[ret], nr_xmit - ret);
-
-		while (likely(dst_vdev->nr_async_pkts)) {
-			nr_cpl = rte_vhost_poll_enqueue_completed(dst_vdev->vid,
-					VIRTIO_RXQ, m_cpl, MAX_PKT_BURST);
-			dst_vdev->nr_async_pkts -= nr_cpl;
-			free_pkts(m_cpl, nr_cpl);
-		}
 	} else {
 		ret = rte_vhost_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ,
 						m, nr_xmit);

@@ -925,7 +925,7 @@ virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 	}
 
 	if (unlikely(vhost_txq->len == MAX_PKT_BURST)) {
-		drain_vhost(dst_vdev, vhost_txq->m_table, MAX_PKT_BURST);
+		drain_vhost(dst_vdev);
 		vhost_txq->len = 0;
 		vhost_tsc[dst_vdev->vid] = rte_rdtsc();
 	}

@@ -1031,7 +1031,7 @@ virtio_tx_route(struct vhost_dev *vdev, struct rte_mbuf *m, uint16_t vlan_tag)
 
 		TAILQ_FOREACH(vdev2, &vhost_dev_list, global_vdev_entry) {
 			if (vdev2 != vdev)
-				virtio_xmit(vdev2, vdev, m);
+				sync_virtio_xmit(vdev2, vdev, m);
 		}
 		goto queue2nic;
 	}

@@ -1124,31 +1124,17 @@ drain_mbuf_table(struct mbuf_table *tx_q)
 	}
 }
 
-static __rte_always_inline void
-complete_async_pkts(struct vhost_dev *vdev, uint16_t qid)
-{
-	struct rte_mbuf *p_cpl[MAX_PKT_BURST];
-	uint16_t complete_count;
-
-	complete_count = rte_vhost_poll_enqueue_completed(vdev->vid,
-						qid, p_cpl, MAX_PKT_BURST);
-	vdev->nr_async_pkts -= complete_count;
-	if (complete_count)
-		free_pkts(p_cpl, complete_count);
-}
-
 static __rte_always_inline void
 drain_eth_rx(struct vhost_dev *vdev)
 {
 	uint16_t rx_count, enqueue_count;
+	uint32_t cpu_cpl_nr;
 	struct rte_mbuf *pkts[MAX_PKT_BURST];
+	struct rte_mbuf *m_cpu_cpl[MAX_PKT_BURST];
 
 	rx_count = rte_eth_rx_burst(ports[0], vdev->vmdq_rx_q,
 				    pkts, MAX_PKT_BURST);
 
-	while (likely(vdev->nr_async_pkts))
-		complete_async_pkts(vdev, VIRTIO_RXQ);
-
 	if (!rx_count)
 		return;
 

@@ -1170,13 +1156,21 @@ drain_eth_rx(struct vhost_dev *vdev)
 		}
 	}
 
+	complete_async_pkts(vdev);
+	while (rte_atomic16_read(&vdev->nr_async_pkts) >= 128)
+		complete_async_pkts(vdev);
+
 	if (builtin_net_driver) {
 		enqueue_count = vs_enqueue_pkts(vdev, VIRTIO_RXQ,
 						pkts, rx_count);
 	} else if (async_vhost_driver) {
 		enqueue_count = rte_vhost_submit_enqueue_burst(vdev->vid,
-					VIRTIO_RXQ, pkts, rx_count);
-		vdev->nr_async_pkts += enqueue_count;
+					VIRTIO_RXQ, pkts, rx_count,
+					m_cpu_cpl, &cpu_cpl_nr);
+		rte_atomic16_add(&vdev->nr_async_pkts,
+					enqueue_count - cpu_cpl_nr);
+		if (cpu_cpl_nr)
+			free_pkts(m_cpu_cpl, cpu_cpl_nr);
 	} else {
 		enqueue_count = rte_vhost_enqueue_burst(vdev->vid, VIRTIO_RXQ,
 						pkts, rx_count);

@@ -1224,7 +1218,7 @@ drain_virtio_tx(struct vhost_dev *vdev)
 			RTE_LOG_DP(DEBUG, VHOST_DATA,
 				"Vhost tX queue drained after timeout with burst size %u\n",
 				vhost_txq->len);
-			drain_vhost(vdev, vhost_txq->m_table, vhost_txq->len);
+			drain_vhost(vdev);
 			vhost_txq->len = 0;
 			vhost_tsc[vdev->vid] = cur_tsc;
 		}

index 4317b6ae81..d33ddb411b 100644
@@ -51,7 +51,7 @@ struct vhost_dev {
 	uint64_t features;
 	size_t hdr_len;
 	uint16_t nr_vrings;
-	uint16_t nr_async_pkts;
+	rte_atomic16_t nr_async_pkts;
 	struct rte_vhost_memory *mem;
 	struct device_statistics stats;
 	TAILQ_ENTRY(vhost_dev) global_vdev_entry;

index 71d8a1f1f5..b0b04aa453 100644
@@ -17,6 +17,7 @@ struct packet_tracker {
 	unsigned short next_read;
 	unsigned short next_write;
 	unsigned short last_remain;
+	unsigned short ioat_space;
 };
 
 struct packet_tracker cb_tracker[MAX_VHOST_DEVICE];

@@ -113,7 +114,7 @@ open_ioat(const char *value)
 			goto out;
 		}
 		rte_rawdev_start(dev_id);
-
+		cb_tracker[dev_id].ioat_space = IOAT_RING_SIZE;
 		dma_info->nr++;
 		i++;
 	}

@@ -140,13 +141,9 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 			src = descs[i_desc].src;
 			dst = descs[i_desc].dst;
 			i_seg = 0;
+			if (cb_tracker[dev_id].ioat_space < src->nr_segs)
+				break;
 			while (i_seg < src->nr_segs) {
-				/*
-				 * TODO: Assuming that the ring space of the
-				 * IOAT device is large enough, so there is no
-				 * error here, and the actual error handling
-				 * will be added later.
-				 */
 				rte_ioat_enqueue_copy(dev_id,
 					(uintptr_t)(src->iov[i_seg].iov_base)
 						+ src->offset,

@@ -158,7 +155,8 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 				i_seg++;
 			}
 			write &= mask;
-			cb_tracker[dev_id].size_track[write] = i_seg;
+			cb_tracker[dev_id].size_track[write] = src->nr_segs;
+			cb_tracker[dev_id].ioat_space -= src->nr_segs;
 			write++;
 		}
 	} else {

@@ -186,6 +184,7 @@ ioat_check_completed_copies_cb(int vid, uint16_t queue_id,
 		int dev_id = dma_bind[vid].dmas[queue_id * 2
 				+ VIRTIO_RXQ].dev_id;
 		n_seg = rte_ioat_completed_ops(dev_id, 255, dump, dump);
+		cb_tracker[dev_id].ioat_space += n_seg;
 		n_seg += cb_tracker[dev_id].last_remain;
 		if (!n_seg)
 			return 0;

index 8d8c3038bf..d400939a2d 100644
@@ -179,9 +179,18 @@ struct mbuf_table {
 	struct rte_mbuf *m_table[MAX_PKT_BURST];
 };
 
+struct vhost_bufftable {
+	uint32_t len;
+	uint64_t pre_tsc;
+	struct rte_mbuf *m_table[MAX_PKT_BURST];
+};
+
 /* TX queue for each data core. */
 struct mbuf_table lcore_tx_queue[RTE_MAX_LCORE];
 
+/* TX queue for each vhost device. */
+struct vhost_bufftable vhost_bufftable[RTE_MAX_LCORE * MAX_VHOST_DEVICE];
+
 #define MBUF_TABLE_DRAIN_TSC	((rte_get_tsc_hz() + US_PER_S - 1) \
 				 / US_PER_S * BURST_TX_DRAIN_US)
 #define VLAN_HLEN       4

@@ -804,39 +813,82 @@ unlink_vmdq(struct vhost_dev *vdev)
 	}
 }
 
+static inline void
+free_pkts(struct rte_mbuf **pkts, uint16_t n)
+{
+	while (n--)
+		rte_pktmbuf_free(pkts[n]);
+}
+
 static __rte_always_inline void
-virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
+complete_async_pkts(struct vhost_dev *vdev)
+{
+	struct rte_mbuf *p_cpl[MAX_PKT_BURST];
+	uint16_t complete_count;
+
+	complete_count = rte_vhost_poll_enqueue_completed(vdev->vid,
+					VIRTIO_RXQ, p_cpl, MAX_PKT_BURST);
+	if (complete_count) {
+		atomic_fetch_sub(&vdev->nr_async_pkts, complete_count);
+		free_pkts(p_cpl, complete_count);
+	}
+}
+
+static __rte_always_inline void
+sync_virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
 	    struct rte_mbuf *m)
 {
 	uint16_t ret;
-	struct rte_mbuf *m_cpl[1];
 
 	if (builtin_net_driver) {
 		ret = vs_enqueue_pkts(dst_vdev, VIRTIO_RXQ, &m, 1);
-	} else if (async_vhost_driver) {
-		ret = rte_vhost_submit_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ,
-						&m, 1);
-
-		if (likely(ret))
-			dst_vdev->nr_async_pkts++;
-
-		while (likely(dst_vdev->nr_async_pkts)) {
-			if (rte_vhost_poll_enqueue_completed(dst_vdev->vid,
-					VIRTIO_RXQ, m_cpl, 1))
-				dst_vdev->nr_async_pkts--;
-		}
 	} else {
 		ret = rte_vhost_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ, &m, 1);
 	}
 
 	if (enable_stats) {
-		rte_atomic64_inc(&dst_vdev->stats.rx_total_atomic);
-		rte_atomic64_add(&dst_vdev->stats.rx_atomic, ret);
+		atomic_fetch_add(&dst_vdev->stats.rx_total_atomic, 1);
+		atomic_fetch_add(&dst_vdev->stats.rx_atomic, ret);
 		src_vdev->stats.tx_total++;
 		src_vdev->stats.tx += ret;
 	}
 }
 
+static __rte_always_inline void
+drain_vhost(struct vhost_dev *vdev)
+{
+	uint16_t ret;
+	uint64_t queue_id = rte_lcore_id() * MAX_VHOST_DEVICE + vdev->vid;
+	uint16_t nr_xmit = vhost_bufftable[queue_id].len;
+	struct rte_mbuf **m = vhost_bufftable[queue_id].m_table;
+
+	if (builtin_net_driver) {
+		ret = vs_enqueue_pkts(vdev, VIRTIO_RXQ, m, nr_xmit);
+	} else if (async_vhost_driver) {
+		uint32_t cpu_cpl_nr;
+		struct rte_mbuf *m_cpu_cpl[nr_xmit];
+		complete_async_pkts(vdev);
+		ret = rte_vhost_submit_enqueue_burst(vdev->vid, VIRTIO_RXQ,
+					m, nr_xmit, m_cpu_cpl, &cpu_cpl_nr);
+		atomic_fetch_add(&vdev->nr_async_pkts, ret - cpu_cpl_nr);
+		if (cpu_cpl_nr)
+			free_pkts(m_cpu_cpl, cpu_cpl_nr);
+		if (nr_xmit - ret)
+			free_pkts(&m[ret], nr_xmit - ret);
+	} else {
+		ret = rte_vhost_enqueue_burst(vdev->vid, VIRTIO_RXQ,
+						m, nr_xmit);
+	}
+
+	if (enable_stats) {
+		atomic_fetch_add(&vdev->stats.rx_total_atomic, nr_xmit);
+		atomic_fetch_add(&vdev->stats.rx_atomic, ret);
+	}
+
+	if (!async_vhost_driver)
+		free_pkts(m, nr_xmit);
+}
+
 /*
  * Check if the packet destination MAC address is for a local device. If so then put
  * the packet on that devices RX queue. If not then return.

@@ -846,7 +898,8 @@ virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 {
 	struct rte_ether_hdr *pkt_hdr;
 	struct vhost_dev *dst_vdev;
-
+	struct vhost_bufftable *vhost_txq;
+	const uint16_t lcore_id = rte_lcore_id();
 	pkt_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
 
 	dst_vdev = find_vhost_dev(&pkt_hdr->d_addr);

@@ -869,7 +922,20 @@ virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 		return 0;
 	}
 
-	virtio_xmit(dst_vdev, vdev, m);
+	vhost_txq = &vhost_bufftable[lcore_id * MAX_VHOST_DEVICE
+					+ dst_vdev->vid];
+	vhost_txq->m_table[vhost_txq->len++] = m;
+
+	if (enable_stats) {
+		vdev->stats.tx_total++;
+		vdev->stats.tx++;
+	}
+
+	if (unlikely(vhost_txq->len == MAX_PKT_BURST)) {
+		drain_vhost(dst_vdev);
+		vhost_txq->len = 0;
+		vhost_txq->pre_tsc = rte_rdtsc();
+	}
 	return 0;
 }
 

@@ -940,13 +1006,6 @@ static void virtio_tx_offload(struct rte_mbuf *m)
 	tcp_hdr->cksum = get_psd_sum(l3_hdr, m->ol_flags);
 }
 
-static inline void
-free_pkts(struct rte_mbuf **pkts, uint16_t n)
-{
-	while (n--)
-		rte_pktmbuf_free(pkts[n]);
-}
-
 static __rte_always_inline void
 do_drain_mbuf_table(struct mbuf_table *tx_q)
 {

@@ -979,14 +1038,13 @@ virtio_tx_route(struct vhost_dev *vdev, struct rte_mbuf *m, uint16_t vlan_tag)
 
 		TAILQ_FOREACH(vdev2, &vhost_dev_list, global_vdev_entry) {
 			if (vdev2 != vdev)
-				virtio_xmit(vdev2, vdev, m);
+				sync_virtio_xmit(vdev2, vdev, m);
 		}
 		goto queue2nic;
 	}
 
 	/*check if destination is local VM*/
 	if ((vm2vm_mode == VM2VM_SOFTWARE) && (virtio_tx_local(vdev, m) == 0)) {
-		rte_pktmbuf_free(m);
 		return;
 	}
 

@@ -1073,19 +1131,6 @@ drain_mbuf_table(struct mbuf_table *tx_q)
 	}
 }
 
-static __rte_always_inline void
-complete_async_pkts(struct vhost_dev *vdev, uint16_t qid)
-{
-	struct rte_mbuf *p_cpl[MAX_PKT_BURST];
-	uint16_t complete_count;
-
-	complete_count = rte_vhost_poll_enqueue_completed(vdev->vid,
-						qid, p_cpl, MAX_PKT_BURST);
-	vdev->nr_async_pkts -= complete_count;
-	if (complete_count)
-		free_pkts(p_cpl, complete_count);
-}
-
 static __rte_always_inline void
 drain_eth_rx(struct vhost_dev *vdev)
 {

@@ -1095,9 +1140,6 @@ drain_eth_rx(struct vhost_dev *vdev)
 	rx_count = rte_eth_rx_burst(ports[0], vdev->vmdq_rx_q,
 				    pkts, MAX_PKT_BURST);
 
-	while (likely(vdev->nr_async_pkts))
-		complete_async_pkts(vdev, VIRTIO_RXQ);
-
 	if (!rx_count)
 		return;
 

@@ -1123,17 +1165,26 @@ drain_eth_rx(struct vhost_dev *vdev)
 		enqueue_count = vs_enqueue_pkts(vdev, VIRTIO_RXQ,
 						pkts, rx_count);
 	} else if (async_vhost_driver) {
+		uint32_t cpu_cpl_nr;
+		struct rte_mbuf *m_cpu_cpl[MAX_PKT_BURST];
+		complete_async_pkts(vdev);
 		enqueue_count = rte_vhost_submit_enqueue_burst(vdev->vid,
-					VIRTIO_RXQ, pkts, rx_count);
-		vdev->nr_async_pkts += enqueue_count;
+					VIRTIO_RXQ, pkts, rx_count,
+					m_cpu_cpl, &cpu_cpl_nr);
+		atomic_fetch_add(&vdev->nr_async_pkts,
+					enqueue_count - cpu_cpl_nr);
+		if (enqueue_count > cpu_cpl_nr)
+			free_pkts(m_cpu_cpl, cpu_cpl_nr);
+		free_pkts(&pkts[enqueue_count], rx_count - enqueue_count);
+
 	} else {
 		enqueue_count = rte_vhost_enqueue_burst(vdev->vid, VIRTIO_RXQ,
 						pkts, rx_count);
 	}
 
 	if (enable_stats) {
-		rte_atomic64_add(&vdev->stats.rx_total_atomic, rx_count);
-		rte_atomic64_add(&vdev->stats.rx_atomic, enqueue_count);
+		atomic_fetch_add(&vdev->stats.rx_total_atomic, rx_count);
+		atomic_fetch_add(&vdev->stats.rx_atomic, enqueue_count);
 	}
 
 	if (!async_vhost_driver)

@@ -1144,8 +1195,11 @@ static __rte_always_inline void
 drain_virtio_tx(struct vhost_dev *vdev)
 {
 	struct rte_mbuf *pkts[MAX_PKT_BURST];
+	struct vhost_bufftable *vhost_txq;
+	const uint16_t lcore_id = rte_lcore_id();
 	uint16_t count;
 	uint16_t i;
+	uint64_t cur_tsc;
 
 	if (builtin_net_driver) {
 		count = vs_dequeue_pkts(vdev, VIRTIO_TXQ, mbuf_pool,

@@ -1163,6 +1217,17 @@ drain_virtio_tx(struct vhost_dev *vdev)
 
 	for (i = 0; i < count; ++i)
 		virtio_tx_route(vdev, pkts[i], vlan_tags[vdev->vid]);
+
+	vhost_txq = &vhost_bufftable[lcore_id * MAX_VHOST_DEVICE + vdev->vid];
+	cur_tsc = rte_rdtsc();
+	if (unlikely(cur_tsc - vhost_txq->pre_tsc > MBUF_TABLE_DRAIN_TSC)) {
+		RTE_LOG_DP(DEBUG, VHOST_DATA,
+			"Vhost tX queue drained after timeout with burst size %u\n",
+			vhost_txq->len);
+		drain_vhost(vdev);
+		vhost_txq->len = 0;
+		vhost_txq->pre_tsc = cur_tsc;
+	}
 }
 
 /*

@@ -1392,8 +1457,8 @@ print_stats(__rte_unused void *arg)
 			tx         = vdev->stats.tx;
 			tx_dropped = tx_total - tx;
 
-			rx_total   = rte_atomic64_read(&vdev->stats.rx_total_atomic);
-			rx         = rte_atomic64_read(&vdev->stats.rx_atomic);
+			rx_total   = atomic_load(&vdev->stats.rx_total_atomic);
+			rx         = atomic_load(&vdev->stats.rx_atomic);
 			rx_dropped = rx_total - rx;
 
 			printf("Statistics for device %d\n"

index 4317b6ae81..6aa798a3e2 100644
@@ -8,6 +8,7 @@
 #include <sys/queue.h>
 
 #include <rte_ether.h>
+#include <stdatomic.h>
 
 /* Macros for printing using RTE_LOG */
 #define RTE_LOGTYPE_VHOST_CONFIG RTE_LOGTYPE_USER1

@@ -21,8 +22,8 @@ enum {VIRTIO_RXQ, VIRTIO_TXQ, VIRTIO_QNUM};
 struct device_statistics {
 	uint64_t	tx;
 	uint64_t	tx_total;
-	rte_atomic64_t	rx_atomic;
-	rte_atomic64_t	rx_total_atomic;
+	atomic_int_least64_t	rx_atomic;
+	atomic_int_least64_t	rx_total_atomic;
 };
 
 struct vhost_queue {

@@ -51,7 +52,7 @@ struct vhost_dev {
 	uint64_t features;
 	size_t hdr_len;
 	uint16_t nr_vrings;
-	uint16_t nr_async_pkts;
+	atomic_int_least16_t nr_async_pkts;
 	struct rte_vhost_memory *mem;
 	struct device_statistics stats;
 	TAILQ_ENTRY(vhost_dev) global_vdev_entry;

index 71d8a1f1f..b0b04aa45 100644
@@ -17,6 +17,7 @@ struct packet_tracker {
 	unsigned short next_read;
 	unsigned short next_write;
 	unsigned short last_remain;
+	unsigned short ioat_space;
 };
 
 struct packet_tracker cb_tracker[MAX_VHOST_DEVICE];

@@ -113,7 +114,7 @@ open_ioat(const char *value)
 			goto out;
 		}
 		rte_rawdev_start(dev_id);
-
+		cb_tracker[dev_id].ioat_space = IOAT_RING_SIZE;
 		dma_info->nr++;
 		i++;
 	}

@@ -140,13 +141,9 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 			src = descs[i_desc].src;
 			dst = descs[i_desc].dst;
 			i_seg = 0;
+			if (cb_tracker[dev_id].ioat_space < src->nr_segs)
+				break;
 			while (i_seg < src->nr_segs) {
-				/*
-				 * TODO: Assuming that the ring space of the
-				 * IOAT device is large enough, so there is no
-				 * error here, and the actual error handling
-				 * will be added later.
-				 */
 				rte_ioat_enqueue_copy(dev_id,
 					(uintptr_t)(src->iov[i_seg].iov_base)
 						+ src->offset,

@@ -158,7 +155,8 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 				i_seg++;
 			}
 			write &= mask;
-			cb_tracker[dev_id].size_track[write] = i_seg;
+			cb_tracker[dev_id].size_track[write] = src->nr_segs;
+			cb_tracker[dev_id].ioat_space -= src->nr_segs;
 			write++;
 		}
 	} else {

@@ -186,6 +184,7 @@ ioat_check_completed_copies_cb(int vid, uint16_t queue_id,
 		int dev_id = dma_bind[vid].dmas[queue_id * 2
 				+ VIRTIO_RXQ].dev_id;
 		n_seg = rte_ioat_completed_ops(dev_id, 255, dump, dump);
+		cb_tracker[dev_id].ioat_space += n_seg;
 		n_seg += cb_tracker[dev_id].last_remain;
 		if (!n_seg)
 			return 0;

index 8d8c3038b..efc044b28 100644
@@ -179,9 +179,18 @@ struct mbuf_table {
 	struct rte_mbuf *m_table[MAX_PKT_BURST];
 };
 
+struct vhost_bufftable {
+	uint32_t len;
+	uint64_t pre_tsc;
+	struct rte_mbuf *m_table[MAX_PKT_BURST];
+};
+
 /* TX queue for each data core. */
 struct mbuf_table lcore_tx_queue[RTE_MAX_LCORE];
 
+/* TX queue for each vhost device. */
+struct vhost_bufftable vhost_bufftable[RTE_MAX_LCORE * MAX_VHOST_DEVICE];
+
 #define MBUF_TABLE_DRAIN_TSC	((rte_get_tsc_hz() + US_PER_S - 1) \
 				 / US_PER_S * BURST_TX_DRAIN_US)
 #define VLAN_HLEN       4

@@ -804,39 +813,84 @@ unlink_vmdq(struct vhost_dev *vdev)
 	}
 }
 
+static inline void
+free_pkts(struct rte_mbuf **pkts, uint16_t n)
+{
+	while (n--)
+		rte_pktmbuf_free(pkts[n]);
+}
+
 static __rte_always_inline void
-virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
+complete_async_pkts(struct vhost_dev *vdev)
+{
+	struct rte_mbuf *p_cpl[MAX_PKT_BURST];
+	uint16_t complete_count;
+
+	complete_count = rte_vhost_poll_enqueue_completed(vdev->vid,
+					VIRTIO_RXQ, p_cpl, MAX_PKT_BURST);
+	if (complete_count) {
+		atomic_fetch_sub(&vdev->nr_async_pkts, complete_count);
+		free_pkts(p_cpl, complete_count);
+	}
+}
+
+static __rte_always_inline void
+sync_virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
 	    struct rte_mbuf *m)
 {
 	uint16_t ret;
-	struct rte_mbuf *m_cpl[1];
 
 	if (builtin_net_driver) {
 		ret = vs_enqueue_pkts(dst_vdev, VIRTIO_RXQ, &m, 1);
-	} else if (async_vhost_driver) {
-		ret = rte_vhost_submit_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ,
-						&m, 1);
-
-		if (likely(ret))
-			dst_vdev->nr_async_pkts++;
-
-		while (likely(dst_vdev->nr_async_pkts)) {
-			if (rte_vhost_poll_enqueue_completed(dst_vdev->vid,
-					VIRTIO_RXQ, m_cpl, 1))
-				dst_vdev->nr_async_pkts--;
-		}
 	} else {
 		ret = rte_vhost_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ, &m, 1);
 	}
 
 	if (enable_stats) {
-		rte_atomic64_inc(&dst_vdev->stats.rx_total_atomic);
-		rte_atomic64_add(&dst_vdev->stats.rx_atomic, ret);
+		atomic_fetch_add(&dst_vdev->stats.rx_total_atomic, 1);
+		atomic_fetch_add(&dst_vdev->stats.rx_atomic, ret);
 		src_vdev->stats.tx_total++;
 		src_vdev->stats.tx += ret;
 	}
 }
 
+static __rte_always_inline void
+drain_vhost(struct vhost_dev *vdev)
+{
+	uint16_t ret;
+	uint64_t queue_id = rte_lcore_id() * MAX_VHOST_DEVICE + vdev->vid;
+	uint16_t nr_xmit = vhost_bufftable[queue_id].len;
+	struct rte_mbuf **m = vhost_bufftable[queue_id].m_table;
+
+	if (builtin_net_driver) {
+		ret = vs_enqueue_pkts(vdev, VIRTIO_RXQ, m, nr_xmit);
+	} else if (async_vhost_driver) {
+		uint32_t cpu_cpl_nr = 0;
+		uint16_t enqueue_fail = 0;
+		struct rte_mbuf *m_cpu_cpl[nr_xmit];
+		complete_async_pkts(vdev);
+		ret = rte_vhost_submit_enqueue_burst(vdev->vid, VIRTIO_RXQ,
+					m, nr_xmit, m_cpu_cpl, &cpu_cpl_nr);
+		atomic_fetch_add(&vdev->nr_async_pkts, ret - cpu_cpl_nr);
+		if (cpu_cpl_nr)
+			free_pkts(m_cpu_cpl, cpu_cpl_nr);
+		enqueue_fail = nr_xmit - ret;
+		if (enqueue_fail)
+			free_pkts(&m[ret], nr_xmit - ret);
+	} else {
+		ret = rte_vhost_enqueue_burst(vdev->vid, VIRTIO_RXQ,
+						m, nr_xmit);
+	}
+
+	if (enable_stats) {
+		atomic_fetch_add(&vdev->stats.rx_total_atomic, nr_xmit);
+		atomic_fetch_add(&vdev->stats.rx_atomic, ret);
+	}
+
+	if (!async_vhost_driver)
+		free_pkts(m, nr_xmit);
+}
+
 /*
  * Check if the packet destination MAC address is for a local device. If so then put
  * the packet on that devices RX queue. If not then return.

@@ -846,7 +900,8 @@ virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 {
 	struct rte_ether_hdr *pkt_hdr;
 	struct vhost_dev *dst_vdev;
-
+	struct vhost_bufftable *vhost_txq;
+	const uint16_t lcore_id = rte_lcore_id();
 	pkt_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
 
 	dst_vdev = find_vhost_dev(&pkt_hdr->d_addr);

@@ -869,7 +924,20 @@ virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 		return 0;
 	}
 
-	virtio_xmit(dst_vdev, vdev, m);
+	vhost_txq = &vhost_bufftable[lcore_id * MAX_VHOST_DEVICE
+					+ dst_vdev->vid];
+	vhost_txq->m_table[vhost_txq->len++] = m;
+
+	if (enable_stats) {
+		vdev->stats.tx_total++;
+		vdev->stats.tx++;
+	}
+
+	if (unlikely(vhost_txq->len == MAX_PKT_BURST)) {
+		drain_vhost(dst_vdev);
+		vhost_txq->len = 0;
+		vhost_txq->pre_tsc = rte_rdtsc();
+	}
 	return 0;
 }
 

@@ -940,13 +1008,6 @@ static void virtio_tx_offload(struct rte_mbuf *m)
 	tcp_hdr->cksum = get_psd_sum(l3_hdr, m->ol_flags);
 }
 
-static inline void
-free_pkts(struct rte_mbuf **pkts, uint16_t n)
-{
-	while (n--)
-		rte_pktmbuf_free(pkts[n]);
-}
-
 static __rte_always_inline void
 do_drain_mbuf_table(struct mbuf_table *tx_q)
 {

@@ -979,14 +1040,13 @@ virtio_tx_route(struct vhost_dev *vdev, struct rte_mbuf *m, uint16_t vlan_tag)
 
 		TAILQ_FOREACH(vdev2, &vhost_dev_list, global_vdev_entry) {
 			if (vdev2 != vdev)
-				virtio_xmit(vdev2, vdev, m);
+				sync_virtio_xmit(vdev2, vdev, m);
 		}
 		goto queue2nic;
 	}
 
 	/*check if destination is local VM*/
 	if ((vm2vm_mode == VM2VM_SOFTWARE) && (virtio_tx_local(vdev, m) == 0)) {
-		rte_pktmbuf_free(m);
 		return;
 	}
 

@@ -1073,19 +1133,6 @@ drain_mbuf_table(struct mbuf_table *tx_q)
 	}
 }
 
-static __rte_always_inline void
-complete_async_pkts(struct vhost_dev *vdev, uint16_t qid)
-{
-	struct rte_mbuf *p_cpl[MAX_PKT_BURST];
-	uint16_t complete_count;
-
-	complete_count = rte_vhost_poll_enqueue_completed(vdev->vid,
-						qid, p_cpl, MAX_PKT_BURST);
-	vdev->nr_async_pkts -= complete_count;
-	if (complete_count)
-		free_pkts(p_cpl, complete_count);
-}
-
 static __rte_always_inline void
 drain_eth_rx(struct vhost_dev *vdev)
 {

@@ -1095,9 +1142,6 @@ drain_eth_rx(struct vhost_dev *vdev)
 	rx_count = rte_eth_rx_burst(ports[0], vdev->vmdq_rx_q,
 				    pkts, MAX_PKT_BURST);
 
-	while (likely(vdev->nr_async_pkts))
-		complete_async_pkts(vdev, VIRTIO_RXQ);
-
 	if (!rx_count)
 		return;
 

@@ -1123,17 +1167,29 @@ drain_eth_rx(struct vhost_dev *vdev)
 		enqueue_count = vs_enqueue_pkts(vdev, VIRTIO_RXQ,
 						pkts, rx_count);
 	} else if (async_vhost_driver) {
+		uint32_t cpu_cpl_nr = 0;
+		uint16_t enqueue_fail = 0;
+		struct rte_mbuf *m_cpu_cpl[MAX_PKT_BURST];
+		complete_async_pkts(vdev);
 		enqueue_count = rte_vhost_submit_enqueue_burst(vdev->vid,
-					VIRTIO_RXQ, pkts, rx_count);
-		vdev->nr_async_pkts += enqueue_count;
+					VIRTIO_RXQ, pkts, rx_count,
+					m_cpu_cpl, &cpu_cpl_nr);
+		atomic_fetch_add(&vdev->nr_async_pkts,
+					enqueue_count - cpu_cpl_nr);
+		if (cpu_cpl_nr)
+			free_pkts(m_cpu_cpl, cpu_cpl_nr);
+		enqueue_fail = rx_count - enqueue_count;
+		if (enqueue_fail)
+			free_pkts(&pkts[enqueue_count], enqueue_fail);
+
 	} else {
 		enqueue_count = rte_vhost_enqueue_burst(vdev->vid, VIRTIO_RXQ,
 						pkts, rx_count);
 	}
 
 	if (enable_stats) {
-		rte_atomic64_add(&vdev->stats.rx_total_atomic, rx_count);
-		rte_atomic64_add(&vdev->stats.rx_atomic, enqueue_count);
+		atomic_fetch_add(&vdev->stats.rx_total_atomic, rx_count);
+		atomic_fetch_add(&vdev->stats.rx_atomic, enqueue_count);
 	}
 
 	if (!async_vhost_driver)

@@ -1144,8 +1200,11 @@ static __rte_always_inline void
 drain_virtio_tx(struct vhost_dev *vdev)
 {
 	struct rte_mbuf *pkts[MAX_PKT_BURST];
+	struct vhost_bufftable *vhost_txq;
+	const uint16_t lcore_id = rte_lcore_id();
 	uint16_t count;
 	uint16_t i;
+	uint64_t cur_tsc;
 
 	if (builtin_net_driver) {
 		count = vs_dequeue_pkts(vdev, VIRTIO_TXQ, mbuf_pool,

@@ -1163,6 +1222,17 @@ drain_virtio_tx(struct vhost_dev *vdev)
 
 	for (i = 0; i < count; ++i)
 		virtio_tx_route(vdev, pkts[i], vlan_tags[vdev->vid]);
+
+	vhost_txq = &vhost_bufftable[lcore_id * MAX_VHOST_DEVICE + vdev->vid];
+	cur_tsc = rte_rdtsc();
+	if (unlikely(cur_tsc - vhost_txq->pre_tsc > MBUF_TABLE_DRAIN_TSC)) {
+		RTE_LOG_DP(DEBUG, VHOST_DATA,
+			"Vhost tX queue drained after timeout with burst size %u\n",
+			vhost_txq->len);
+		drain_vhost(vdev);
+		vhost_txq->len = 0;
+		vhost_txq->pre_tsc = cur_tsc;
+	}
 }
 
 /*

@@ -1392,8 +1462,8 @@ print_stats(__rte_unused void *arg)
 			tx         = vdev->stats.tx;
 			tx_dropped = tx_total - tx;
 
-			rx_total   = rte_atomic64_read(&vdev->stats.rx_total_atomic);
-			rx         = rte_atomic64_read(&vdev->stats.rx_atomic);
+			rx_total   = atomic_load(&vdev->stats.rx_total_atomic);
+			rx         = atomic_load(&vdev->stats.rx_atomic);
 			rx_dropped = rx_total - rx;
 
 			printf("Statistics for device %d\n"

index 4317b6ae8..6aa798a3e 100644
@@ -8,6 +8,7 @@
 #include <sys/queue.h>
 
 #include <rte_ether.h>
+#include <stdatomic.h>
 
 /* Macros for printing using RTE_LOG */
 #define RTE_LOGTYPE_VHOST_CONFIG RTE_LOGTYPE_USER1

@@ -21,8 +22,8 @@ enum {VIRTIO_RXQ, VIRTIO_TXQ, VIRTIO_QNUM};
 struct device_statistics {
 	uint64_t	tx;
 	uint64_t	tx_total;
-	rte_atomic64_t	rx_atomic;
-	rte_atomic64_t	rx_total_atomic;
+	atomic_int_least64_t	rx_atomic;
+	atomic_int_least64_t	rx_total_atomic;
 };
 
 struct vhost_queue {

@@ -51,7 +52,7 @@ struct vhost_dev {
 	uint64_t features;
 	size_t hdr_len;
 	uint16_t nr_vrings;
-	uint16_t nr_async_pkts;
+	atomic_int_least16_t nr_async_pkts;
 	struct rte_vhost_memory *mem;
 	struct device_statistics stats;
 	TAILQ_ENTRY(vhost_dev) global_vdev_entry;

index 71d8a1f1f..b0b04aa45 100644
@@ -17,6 +17,7 @@ struct packet_tracker {
 	unsigned short next_read;
 	unsigned short next_write;
 	unsigned short last_remain;
+	unsigned short ioat_space;
 };
 
 struct packet_tracker cb_tracker[MAX_VHOST_DEVICE];

@@ -113,7 +114,7 @@ open_ioat(const char *value)
 			goto out;
 		}
 		rte_rawdev_start(dev_id);
-
+		cb_tracker[dev_id].ioat_space = IOAT_RING_SIZE;
 		dma_info->nr++;
 		i++;
 	}

@@ -140,13 +141,9 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 			src = descs[i_desc].src;
 			dst = descs[i_desc].dst;
 			i_seg = 0;
+			if (cb_tracker[dev_id].ioat_space < src->nr_segs)
+				break;
 			while (i_seg < src->nr_segs) {
-				/*
-				 * TODO: Assuming that the ring space of the
-				 * IOAT device is large enough, so there is no
-				 * error here, and the actual error handling
-				 * will be added later.
-				 */
 				rte_ioat_enqueue_copy(dev_id,
 					(uintptr_t)(src->iov[i_seg].iov_base)
 						+ src->offset,

@@ -158,7 +155,8 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 				i_seg++;
 			}
 			write &= mask;
-			cb_tracker[dev_id].size_track[write] = i_seg;
+			cb_tracker[dev_id].size_track[write] = src->nr_segs;
+			cb_tracker[dev_id].ioat_space -= src->nr_segs;
 			write++;
 		}
 	} else {

@@ -186,6 +184,7 @@ ioat_check_completed_copies_cb(int vid, uint16_t queue_id,
 		int dev_id = dma_bind[vid].dmas[queue_id * 2
 				+ VIRTIO_RXQ].dev_id;
 		n_seg = rte_ioat_completed_ops(dev_id, 255, dump, dump);
+		cb_tracker[dev_id].ioat_space += n_seg;
 		n_seg += cb_tracker[dev_id].last_remain;
 		if (!n_seg)
 			return 0;

index 8d8c3038b..3ea12a474 100644
@@ -179,9 +179,18 @@ struct mbuf_table {
 	struct rte_mbuf *m_table[MAX_PKT_BURST];
 };
 
+struct vhost_bufftable {
+	uint32_t len;
+	uint64_t pre_tsc;
+	struct rte_mbuf *m_table[MAX_PKT_BURST];
+};
+
 /* TX queue for each data core. */
 struct mbuf_table lcore_tx_queue[RTE_MAX_LCORE];
 
+/* TX queue for each vhost device. */
+struct vhost_bufftable *vhost_txbuff[RTE_MAX_LCORE * MAX_VHOST_DEVICE];
+
 #define MBUF_TABLE_DRAIN_TSC	((rte_get_tsc_hz() + US_PER_S - 1) \
 				 / US_PER_S * BURST_TX_DRAIN_US)
 #define VLAN_HLEN       4

@@ -804,39 +813,114 @@ unlink_vmdq(struct vhost_dev *vdev)
 	}
 }
 
+static inline void
+free_pkts(struct rte_mbuf **pkts, uint16_t n)
+{
+	while (n--)
+		rte_pktmbuf_free(pkts[n]);
+}
+
 static __rte_always_inline void
-virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
+complete_async_pkts(struct vhost_dev *vdev)
+{
+	struct rte_mbuf *p_cpl[MAX_PKT_BURST];
+	uint16_t complete_count;
+
+	complete_count = rte_vhost_poll_enqueue_completed(vdev->vid,
+					VIRTIO_RXQ, p_cpl, MAX_PKT_BURST);
+	if (complete_count) {
+		atomic_fetch_sub(&vdev->nr_async_pkts, complete_count);
+		free_pkts(p_cpl, complete_count);
+	}
+}
+
+static __rte_always_inline void
+sync_virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
 	    struct rte_mbuf *m)
 {
 	uint16_t ret;
-	struct rte_mbuf *m_cpl[1];
 
 	if (builtin_net_driver) {
 		ret = vs_enqueue_pkts(dst_vdev, VIRTIO_RXQ, &m, 1);
-	} else if (async_vhost_driver) {
-		ret = rte_vhost_submit_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ,
-						&m, 1);
-
-		if (likely(ret))
-			dst_vdev->nr_async_pkts++;
-
-		while (likely(dst_vdev->nr_async_pkts)) {
-			if (rte_vhost_poll_enqueue_completed(dst_vdev->vid,
-					VIRTIO_RXQ, m_cpl, 1))
-				dst_vdev->nr_async_pkts--;
-		}
 	} else {
 		ret = rte_vhost_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ, &m, 1);
 	}
 
 	if (enable_stats) {
-		rte_atomic64_inc(&dst_vdev->stats.rx_total_atomic);
-		rte_atomic64_add(&dst_vdev->stats.rx_atomic, ret);
+		atomic_fetch_add(&dst_vdev->stats.rx_total_atomic, 1);
+		atomic_fetch_add(&dst_vdev->stats.rx_atomic, ret);
 		src_vdev->stats.tx_total++;
 		src_vdev->stats.tx += ret;
 	}
 }
 
+static __rte_always_inline void
+drain_vhost(struct vhost_dev *vdev)
+{
+	uint16_t ret;
+	uint64_t queue_id = rte_lcore_id() * MAX_VHOST_DEVICE + vdev->vid;
+	uint16_t nr_xmit = vhost_txbuff[queue_id]->len;
+	struct rte_mbuf **m = vhost_txbuff[queue_id]->m_table;
+
+	if (builtin_net_driver) {
+		ret = vs_enqueue_pkts(vdev, VIRTIO_RXQ, m, nr_xmit);
+	} else if (async_vhost_driver) {
+		uint32_t cpu_cpl_nr = 0;
+		uint16_t enqueue_fail = 0;
+		struct rte_mbuf *m_cpu_cpl[nr_xmit];
+
+		complete_async_pkts(vdev);
+		ret = rte_vhost_submit_enqueue_burst(vdev->vid, VIRTIO_RXQ,
+					m, nr_xmit, m_cpu_cpl, &cpu_cpl_nr);
+		atomic_fetch_add(&vdev->nr_async_pkts, ret - cpu_cpl_nr);
+
+		if (cpu_cpl_nr)
+			free_pkts(m_cpu_cpl, cpu_cpl_nr);
+
+		enqueue_fail = nr_xmit - ret;
+		if (enqueue_fail)
+			free_pkts(&m[ret], nr_xmit - ret);
+	} else {
+		ret = rte_vhost_enqueue_burst(vdev->vid, VIRTIO_RXQ,
+						m, nr_xmit);
+	}
+
+	if (enable_stats) {
+		atomic_fetch_add(&vdev->stats.rx_total_atomic, nr_xmit);
+		atomic_fetch_add(&vdev->stats.rx_atomic, ret);
+	}
+
+	if (!async_vhost_driver)
+		free_pkts(m, nr_xmit);
+}
+
+static __rte_always_inline void
+drain_vhost_table(void)
+{
+	const uint16_t lcore_id = rte_lcore_id();
+	struct vhost_bufftable *vhost_txq;
+	struct vhost_dev *vdev;
+	uint64_t cur_tsc;
+
+	TAILQ_FOREACH(vdev, &lcore_info[lcore_id].vdev_list, lcore_vdev_entry) {
+		if (!vdev->remove) {
+			vhost_txq = vhost_txbuff[lcore_id * MAX_VHOST_DEVICE
+					+ vdev->vid];
+			cur_tsc = rte_rdtsc();
+
+			if (unlikely(cur_tsc - vhost_txq->pre_tsc
+					> MBUF_TABLE_DRAIN_TSC)) {
+				RTE_LOG_DP(DEBUG, VHOST_DATA,
+					"Vhost tX queue drained after timeout with burst size %u\n",
+					vhost_txq->len);
+				drain_vhost(vdev);
+				vhost_txq->len = 0;
+				vhost_txq->pre_tsc = cur_tsc;
+			}
+		}
+	}
+}
+
 /*
  * Check if the packet destination MAC address is for a local device. If so then put
  * the packet on that devices RX queue. If not then return.

@@ -846,7 +930,8 @@ virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 {
 	struct rte_ether_hdr *pkt_hdr;
 	struct vhost_dev *dst_vdev;
-
+	struct vhost_bufftable *vhost_txq;
+	const uint16_t lcore_id = rte_lcore_id();
 	pkt_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
 
 	dst_vdev = find_vhost_dev(&pkt_hdr->d_addr);

@@ -869,7 +954,19 @@ virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 		return 0;
 	}
 
-	virtio_xmit(dst_vdev, vdev, m);
+	vhost_txq = vhost_txbuff[lcore_id * MAX_VHOST_DEVICE + dst_vdev->vid];
+	vhost_txq->m_table[vhost_txq->len++] = m;
+
+	if (enable_stats) {
+		vdev->stats.tx_total++;
+		vdev->stats.tx++;
+	}
+
+	if (unlikely(vhost_txq->len == MAX_PKT_BURST)) {
+		drain_vhost(dst_vdev);
+		vhost_txq->len = 0;
+		vhost_txq->pre_tsc = rte_rdtsc();
+	}
 	return 0;
 }
 

@@ -940,13 +1037,6 @@ static void virtio_tx_offload(struct rte_mbuf *m)
 	tcp_hdr->cksum = get_psd_sum(l3_hdr, m->ol_flags);
 }
 
-static inline void
-free_pkts(struct rte_mbuf **pkts, uint16_t n)
-{
-	while (n--)
-		rte_pktmbuf_free(pkts[n]);
-}
-
 static __rte_always_inline void
 do_drain_mbuf_table(struct mbuf_table *tx_q)
 {

@@ -979,16 +1069,14 @@ virtio_tx_route(struct vhost_dev *vdev, struct rte_mbuf *m, uint16_t vlan_tag)
 
 		TAILQ_FOREACH(vdev2, &vhost_dev_list, global_vdev_entry) {
 			if (vdev2 != vdev)
-				virtio_xmit(vdev2, vdev, m);
+				sync_virtio_xmit(vdev2, vdev, m);
 		}
 		goto queue2nic;
 	}
 
 	/*check if destination is local VM*/
-	if ((vm2vm_mode == VM2VM_SOFTWARE) && (virtio_tx_local(vdev, m) == 0)) {
-		rte_pktmbuf_free(m);
+	if ((vm2vm_mode == VM2VM_SOFTWARE) && (virtio_tx_local(vdev, m) == 0))
 		return;
-	}
 
 	if (unlikely(vm2vm_mode == VM2VM_HARDWARE)) {
 		if (unlikely(find_local_dest(vdev, m, &offset,

@@ -1073,19 +1161,6 @@ drain_mbuf_table(struct mbuf_table *tx_q)
 	}
 }
 
-static __rte_always_inline void
-complete_async_pkts(struct vhost_dev *vdev, uint16_t qid)
-{
-	struct rte_mbuf *p_cpl[MAX_PKT_BURST];
-	uint16_t complete_count;
-
-	complete_count = rte_vhost_poll_enqueue_completed(vdev->vid,
-						qid, p_cpl, MAX_PKT_BURST);
-	vdev->nr_async_pkts -= complete_count;
-	if (complete_count)
-		free_pkts(p_cpl, complete_count);
-}
-
 static __rte_always_inline void
 drain_eth_rx(struct vhost_dev *vdev)
 {

@@ -1095,9 +1170,6 @@ drain_eth_rx(struct vhost_dev *vdev)
 	rx_count = rte_eth_rx_burst(ports[0], vdev->vmdq_rx_q,
 				    pkts, MAX_PKT_BURST);
 
-	while (likely(vdev->nr_async_pkts))
-		complete_async_pkts(vdev, VIRTIO_RXQ);
-
 	if (!rx_count)
 		return;
 

@@ -1123,17 +1195,31 @@ drain_eth_rx(struct vhost_dev *vdev)
 		enqueue_count = vs_enqueue_pkts(vdev, VIRTIO_RXQ,
 						pkts, rx_count);
 	} else if (async_vhost_driver) {
+		uint32_t cpu_cpl_nr = 0;
+		uint16_t enqueue_fail = 0;
+		struct rte_mbuf *m_cpu_cpl[MAX_PKT_BURST];
+
+		complete_async_pkts(vdev);
 		enqueue_count = rte_vhost_submit_enqueue_burst(vdev->vid,
-					VIRTIO_RXQ, pkts, rx_count);
-		vdev->nr_async_pkts += enqueue_count;
+					VIRTIO_RXQ, pkts, rx_count,
+					m_cpu_cpl, &cpu_cpl_nr);
+		atomic_fetch_add(&vdev->nr_async_pkts,
+					enqueue_count - cpu_cpl_nr);
+		if (cpu_cpl_nr)
+			free_pkts(m_cpu_cpl, cpu_cpl_nr);
+
+		enqueue_fail = rx_count - enqueue_count;
+		if (enqueue_fail)
+			free_pkts(&pkts[enqueue_count], enqueue_fail);
+
 	} else {
 		enqueue_count = rte_vhost_enqueue_burst(vdev->vid, VIRTIO_RXQ,
 						pkts, rx_count);
 	}
 
 	if (enable_stats) {
-		rte_atomic64_add(&vdev->stats.rx_total_atomic, rx_count);
-		rte_atomic64_add(&vdev->stats.rx_atomic, enqueue_count);
+		atomic_fetch_add(&vdev->stats.rx_total_atomic, rx_count);
+		atomic_fetch_add(&vdev->stats.rx_atomic, enqueue_count);
 	}
 
 	if (!async_vhost_driver)

@@ -1202,7 +1288,7 @@ switch_worker(void *arg __rte_unused)
 
 	while(1) {
 		drain_mbuf_table(tx_q);
-
+		drain_vhost_table();
 		/*
 		 * Inform the configuration core that we have exited the
 		 * linked list and that no devices are in use if requested.

@@ -1298,6 +1384,7 @@ static int
 new_device(int vid)
 {
 	int lcore, core_add = 0;
+	uint16_t i;
 	uint32_t device_num_min = num_devices;
 	struct vhost_dev *vdev;
 	vdev = rte_zmalloc("vhost device", sizeof(*vdev), RTE_CACHE_LINE_SIZE);

@@ -1309,6 +1396,13 @@ new_device(int vid)
 	}
 	vdev->vid = vid;
 
+	for (i = 0; i < RTE_MAX_LCORE; i++) {
+		vhost_txbuff[i * MAX_VHOST_DEVICE + vid]
+			= rte_zmalloc("vhost bufftable",
+				sizeof(struct vhost_bufftable),
+				RTE_CACHE_LINE_SIZE);
+	}
+
 	if (builtin_net_driver)
 		vs_vhost_net_setup(vdev);
 

@@ -1343,12 +1437,15 @@ new_device(int vid)
 	if (async_vhost_driver) {
 		struct rte_vhost_async_features f;
 		struct rte_vhost_async_channel_ops channel_ops;
+
 		if (strncmp(dma_type, "ioat", 4) == 0) {
 			channel_ops.transfer_data = ioat_transfer_data_cb;
 			channel_ops.check_completed_copies =
 				ioat_check_completed_copies_cb;
+
 			f.async_inorder = 1;
 			f.async_threshold = 256;
+
 			return rte_vhost_async_channel_register(vid, VIRTIO_RXQ,
 				f.intval, &channel_ops);
 		}

@@ -1392,8 +1489,8 @@ print_stats(__rte_unused void *arg)
 			tx         = vdev->stats.tx;
 			tx_dropped = tx_total - tx;
 
-			rx_total   = rte_atomic64_read(&vdev->stats.rx_total_atomic);
-			rx         = rte_atomic64_read(&vdev->stats.rx_atomic);
+			rx_total   = atomic_load(&vdev->stats.rx_total_atomic);
+			rx         = atomic_load(&vdev->stats.rx_atomic);
 			rx_dropped = rx_total - rx;
 
 			printf("Statistics for device %d\n"

@@ -1592,6 +1689,7 @@ main(int argc, char *argv[])
 	/* Register vhost user driver to handle vhost messages. */
 	for (i = 0; i < nb_sockets; i++) {
 		char *file = socket_files + i * PATH_MAX;
+
 		if (async_vhost_driver)
 			flags = flags | RTE_VHOST_USER_ASYNC_COPY;
 

index 4317b6ae8..6aa798a3e 100644
@@ -8,6 +8,7 @@
 #include <sys/queue.h>
 
 #include <rte_ether.h>
+#include <stdatomic.h>
 
 /* Macros for printing using RTE_LOG */
 #define RTE_LOGTYPE_VHOST_CONFIG RTE_LOGTYPE_USER1

@@ -21,8 +22,8 @@ enum {VIRTIO_RXQ, VIRTIO_TXQ, VIRTIO_QNUM};
 struct device_statistics {
 	uint64_t	tx;
 	uint64_t	tx_total;
-	rte_atomic64_t	rx_atomic;
-	rte_atomic64_t	rx_total_atomic;
+	atomic_int_least64_t	rx_atomic;
+	atomic_int_least64_t	rx_total_atomic;
 };
 
 struct vhost_queue {

@@ -51,7 +52,7 @@ struct vhost_dev {
 	uint64_t features;
 	size_t hdr_len;
 	uint16_t nr_vrings;
-	uint16_t nr_async_pkts;
+	atomic_int_least16_t nr_async_pkts;
 	struct rte_vhost_memory *mem;
 	struct device_statistics stats;
 	TAILQ_ENTRY(vhost_dev) global_vdev_entry;

index 71d8a1f1f..679d1e2f5 100644
@@ -17,6 +17,7 @@ struct packet_tracker {
 	unsigned short next_read;
 	unsigned short next_write;
 	unsigned short last_remain;
+	unsigned short ioat_space;
 };
 
 struct packet_tracker cb_tracker[MAX_VHOST_DEVICE];

@@ -113,7 +114,7 @@ open_ioat(const char *value)
 			goto out;
 		}
 		rte_rawdev_start(dev_id);
-
+		cb_tracker[dev_id].ioat_space = IOAT_RING_SIZE;
 		dma_info->nr++;
 		i++;
 	}

@@ -140,13 +141,9 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 			src = descs[i_desc].src;
 			dst = descs[i_desc].dst;
 			i_seg = 0;
+			if (cb_tracker[dev_id].ioat_space < src->nr_segs)
+				break;
 			while (i_seg < src->nr_segs) {
-				/*
-				 * TODO: Assuming that the ring space of the
-				 * IOAT device is large enough, so there is no
-				 * error here, and the actual error handling
-				 * will be added later.
-				 */
 				rte_ioat_enqueue_copy(dev_id,
 					(uintptr_t)(src->iov[i_seg].iov_base)
 						+ src->offset,

@@ -158,7 +155,8 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 				i_seg++;
 			}
 			write &= mask;
-			cb_tracker[dev_id].size_track[write] = i_seg;
+			cb_tracker[dev_id].size_track[write] = src->nr_segs;
+			cb_tracker[dev_id].ioat_space -= src->nr_segs;
 			write++;
 		}
 	} else {

@@ -178,17 +176,21 @@ ioat_check_completed_copies_cb(int vid, uint16_t queue_id,
 {
 	if (!opaque_data) {
 		uintptr_t dump[255];
-		unsigned short n_seg;
+		int n_seg;
 		unsigned short read, write;
 		unsigned short nb_packet = 0;
 		unsigned short mask = MAX_ENQUEUED_SIZE - 1;
 		unsigned short i;
+
 		int dev_id = dma_bind[vid].dmas[queue_id * 2
 				+ VIRTIO_RXQ].dev_id;
 		n_seg = rte_ioat_completed_ops(dev_id, 255, dump, dump);
-		n_seg += cb_tracker[dev_id].last_remain;
 		if (!n_seg)
 			return 0;
+
+		cb_tracker[dev_id].ioat_space += n_seg;
+		n_seg += cb_tracker[dev_id].last_remain;
+
 		read = cb_tracker[dev_id].next_read;
 		write = cb_tracker[dev_id].next_write;
 		for (i = 0; i < max_packets; i++) {

index 8d8c3038b..45976c93c 100644
@@ -179,9 +179,22 @@ struct mbuf_table {
 	struct rte_mbuf *m_table[MAX_PKT_BURST];
 };
 
+struct vhost_bufftable {
+	uint32_t len;
+	uint64_t pre_tsc;
+	struct rte_mbuf *m_table[MAX_PKT_BURST];
+};
+
 /* TX queue for each data core. */
 struct mbuf_table lcore_tx_queue[RTE_MAX_LCORE];
 
+/*
+ * Vhost TX buffer for each data core.
+ * Every data core maintains a TX buffer for every vhost device,
+ * which is used for batch pkts enqueue for higher performance.
+ */
+struct vhost_bufftable *vhost_txbuff[RTE_MAX_LCORE * MAX_VHOST_DEVICE];
+
 #define MBUF_TABLE_DRAIN_TSC	((rte_get_tsc_hz() + US_PER_S - 1) \
 				 / US_PER_S * BURST_TX_DRAIN_US)
 #define VLAN_HLEN       4

@@ -804,39 +817,112 @@ unlink_vmdq(struct vhost_dev *vdev)
 	}
 }
 
+static inline void
+free_pkts(struct rte_mbuf **pkts, uint16_t n)
+{
+	while (n--)
+		rte_pktmbuf_free(pkts[n]);
+}
+
 static __rte_always_inline void
-virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
+complete_async_pkts(struct vhost_dev *vdev)
+{
+	struct rte_mbuf *p_cpl[MAX_PKT_BURST];
+	uint16_t complete_count;
+
+	complete_count = rte_vhost_poll_enqueue_completed(vdev->vid,
+					VIRTIO_RXQ, p_cpl, MAX_PKT_BURST);
+	if (complete_count) {
+		atomic_fetch_sub(&vdev->nr_async_pkts, complete_count);
+		free_pkts(p_cpl, complete_count);
+	}
+}
+
+static __rte_always_inline void
+sync_virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
 	    struct rte_mbuf *m)
 {
 	uint16_t ret;
-	struct rte_mbuf *m_cpl[1];
 
 	if (builtin_net_driver) {
 		ret = vs_enqueue_pkts(dst_vdev, VIRTIO_RXQ, &m, 1);
-	} else if (async_vhost_driver) {
-		ret = rte_vhost_submit_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ,
-						&m, 1);
-
-		if (likely(ret))
-			dst_vdev->nr_async_pkts++;
-
-		while (likely(dst_vdev->nr_async_pkts)) {
-			if (rte_vhost_poll_enqueue_completed(dst_vdev->vid,
-					VIRTIO_RXQ, m_cpl, 1))
-				dst_vdev->nr_async_pkts--;
-		}
 	} else {
 		ret = rte_vhost_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ, &m, 1);
 	}
 
 	if (enable_stats) {
-		rte_atomic64_inc(&dst_vdev->stats.rx_total_atomic);
-		rte_atomic64_add(&dst_vdev->stats.rx_atomic, ret);
+		atomic_fetch_add(&dst_vdev->stats.rx_total_atomic, 1);
+		atomic_fetch_add(&dst_vdev->stats.rx_atomic, ret);
 		src_vdev->stats.tx_total++;
 		src_vdev->stats.tx += ret;
 	}
 }
 
+static __rte_always_inline void
+drain_vhost(struct vhost_dev *vdev)
+{
+	uint16_t ret;
+	uint64_t buff_idx = rte_lcore_id() * MAX_VHOST_DEVICE + vdev->vid;
+	uint16_t nr_xmit = vhost_txbuff[buff_idx]->len;
+	struct rte_mbuf **m = vhost_txbuff[buff_idx]->m_table;
+
+	if (builtin_net_driver) {
+		ret = vs_enqueue_pkts(vdev, VIRTIO_RXQ, m, nr_xmit);
+	} else if (async_vhost_driver) {
+		uint32_t cpu_cpl_nr = 0;
+		uint16_t enqueue_fail = 0;
+		struct rte_mbuf *m_cpu_cpl[nr_xmit];
+
+		complete_async_pkts(vdev);
+		ret = rte_vhost_submit_enqueue_burst(vdev->vid, VIRTIO_RXQ,
+					m, nr_xmit, m_cpu_cpl, &cpu_cpl_nr);
+		atomic_fetch_add(&vdev->nr_async_pkts, ret - cpu_cpl_nr);
+
+		if (cpu_cpl_nr)
+			free_pkts(m_cpu_cpl, cpu_cpl_nr);
+
+		enqueue_fail = nr_xmit - ret;
+		if (enqueue_fail)
+			free_pkts(&m[ret], nr_xmit - ret);
+	} else {
+		ret = rte_vhost_enqueue_burst(vdev->vid, VIRTIO_RXQ,
+						m, nr_xmit);
+	}
+
+	if (enable_stats) {
+		atomic_fetch_add(&vdev->stats.rx_total_atomic, nr_xmit);
+		atomic_fetch_add(&vdev->stats.rx_atomic, ret);
+	}
+
+	if (!async_vhost_driver)
+		free_pkts(m, nr_xmit);
+}
+
+static __rte_always_inline void
+drain_vhost_table(void)
+{
+	uint16_t lcore_id = rte_lcore_id();
+	struct vhost_bufftable *vhost_txq;
+	struct vhost_dev *vdev;
+	uint64_t cur_tsc;
+
+	TAILQ_FOREACH(vdev, &vhost_dev_list, global_vdev_entry) {
+		vhost_txq = vhost_txbuff[lcore_id * MAX_VHOST_DEVICE
+						+ vdev->vid];
+
+		cur_tsc = rte_rdtsc();
+		if (unlikely(cur_tsc - vhost_txq->pre_tsc
+				> MBUF_TABLE_DRAIN_TSC)) {
+			RTE_LOG_DP(DEBUG, VHOST_DATA,
+				"Vhost TX queue drained after timeout with burst size %u\n",
+				vhost_txq->len);
+			drain_vhost(vdev);
+			vhost_txq->len = 0;
+			vhost_txq->pre_tsc = cur_tsc;
+		}
+	}
+}
+
 /*
  * Check if the packet destination MAC address is for a local device. If so then put
  * the packet on that devices RX queue. If not then return.

@@ -846,7 +932,8 @@ virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 {
 	struct rte_ether_hdr *pkt_hdr;
 	struct vhost_dev *dst_vdev;
-
+	struct vhost_bufftable *vhost_txq;
+	uint16_t lcore_id = rte_lcore_id();
 	pkt_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
 
 	dst_vdev = find_vhost_dev(&pkt_hdr->d_addr);

@@ -869,7 +956,19 @@ virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 		return 0;
 	}
 
-	virtio_xmit(dst_vdev, vdev, m);
+	vhost_txq = vhost_txbuff[lcore_id * MAX_VHOST_DEVICE + dst_vdev->vid];
+	vhost_txq->m_table[vhost_txq->len++] = m;
+
+	if (enable_stats) {
+		vdev->stats.tx_total++;
+		vdev->stats.tx++;
+	}
+
+	if (unlikely(vhost_txq->len == MAX_PKT_BURST)) {
+		drain_vhost(dst_vdev);
+		vhost_txq->len = 0;
+		vhost_txq->pre_tsc = rte_rdtsc();
+	}
 	return 0;
 }
 

@@ -940,13 +1039,6 @@ static void virtio_tx_offload(struct rte_mbuf *m)
 	tcp_hdr->cksum = get_psd_sum(l3_hdr, m->ol_flags);
 }
 
-static inline void
-free_pkts(struct rte_mbuf **pkts, uint16_t n)
-{
-	while (n--)
-		rte_pktmbuf_free(pkts[n]);
-}
-
 static __rte_always_inline void
 do_drain_mbuf_table(struct mbuf_table *tx_q)
 {

@@ -979,16 +1071,14 @@ virtio_tx_route(struct vhost_dev *vdev, struct rte_mbuf *m, uint16_t vlan_tag)
 
 		TAILQ_FOREACH(vdev2, &vhost_dev_list, global_vdev_entry) {
 			if (vdev2 != vdev)
-				virtio_xmit(vdev2, vdev, m);
+				sync_virtio_xmit(vdev2, vdev, m);
 		}
 		goto queue2nic;
 	}
 
 	/*check if destination is local VM*/
-	if ((vm2vm_mode == VM2VM_SOFTWARE) && (virtio_tx_local(vdev, m) == 0)) {
-		rte_pktmbuf_free(m);
+	if ((vm2vm_mode == VM2VM_SOFTWARE) && (virtio_tx_local(vdev, m) == 0))
 		return;
-	}
 
 	if (unlikely(vm2vm_mode == VM2VM_HARDWARE)) {
 		if (unlikely(find_local_dest(vdev, m, &offset,

@@ -1073,19 +1163,6 @@ drain_mbuf_table(struct mbuf_table *tx_q)
 	}
 }
 
-static __rte_always_inline void
-complete_async_pkts(struct vhost_dev *vdev, uint16_t qid)
-{
-	struct rte_mbuf *p_cpl[MAX_PKT_BURST];
-	uint16_t complete_count;
-
-	complete_count = rte_vhost_poll_enqueue_completed(vdev->vid,
-						qid, p_cpl, MAX_PKT_BURST);
-	vdev->nr_async_pkts -= complete_count;
-	if (complete_count)
-		free_pkts(p_cpl, complete_count);
-}
-
 static __rte_always_inline void
 drain_eth_rx(struct vhost_dev *vdev)
 {

@@ -1095,9 +1172,6 @@ drain_eth_rx(struct vhost_dev *vdev)
 	rx_count = rte_eth_rx_burst(ports[0], vdev->vmdq_rx_q,
 				    pkts, MAX_PKT_BURST);
 
-	while (likely(vdev->nr_async_pkts))
-		complete_async_pkts(vdev, VIRTIO_RXQ);
-
 	if (!rx_count)
 		return;
 

@@ -1123,17 +1197,31 @@ drain_eth_rx(struct vhost_dev *vdev)
 		enqueue_count = vs_enqueue_pkts(vdev, VIRTIO_RXQ,
 						pkts, rx_count);
 	} else if (async_vhost_driver) {
+		uint32_t cpu_cpl_nr = 0;
+		uint16_t enqueue_fail = 0;
+		struct rte_mbuf *m_cpu_cpl[MAX_PKT_BURST];
+
+		complete_async_pkts(vdev);
 		enqueue_count = rte_vhost_submit_enqueue_burst(vdev->vid,
-					VIRTIO_RXQ, pkts, rx_count);
-		vdev->nr_async_pkts += enqueue_count;
+					VIRTIO_RXQ, pkts, rx_count,
+					m_cpu_cpl, &cpu_cpl_nr);
+		atomic_fetch_add(&vdev->nr_async_pkts,
+					enqueue_count - cpu_cpl_nr);
+		if (cpu_cpl_nr)
+			free_pkts(m_cpu_cpl, cpu_cpl_nr);
+
+		enqueue_fail = rx_count - enqueue_count;
+		if (enqueue_fail)
+			free_pkts(&pkts[enqueue_count], enqueue_fail);
+
 	} else {
 		enqueue_count = rte_vhost_enqueue_burst(vdev->vid, VIRTIO_RXQ,
 						pkts, rx_count);
 	}
 
 	if (enable_stats) {
-		rte_atomic64_add(&vdev->stats.rx_total_atomic, rx_count);
-		rte_atomic64_add(&vdev->stats.rx_atomic, enqueue_count);
+		atomic_fetch_add(&vdev->stats.rx_total_atomic, rx_count);
+		atomic_fetch_add(&vdev->stats.rx_atomic, enqueue_count);
 	}
 
 	if (!async_vhost_driver)

@@ -1202,7 +1290,7 @@ switch_worker(void *arg __rte_unused)
 
 	while(1) {
 		drain_mbuf_table(tx_q);
-
+		drain_vhost_table();
 		/*
 		 * Inform the configuration core that we have exited the
 		 * linked list and that no devices are in use if requested.

@@ -1243,6 +1331,7 @@ destroy_device(int vid)
 {
 	struct vhost_dev *vdev = NULL;
 	int lcore;
+	uint16_t i;
 
 	TAILQ_FOREACH(vdev, &vhost_dev_list, global_vdev_entry) {
 		if (vdev->vid == vid)

@@ -1256,6 +1345,9 @@ destroy_device(int vid)
 		rte_pause();
 	}
 
+	for (i = 0; i < RTE_MAX_LCORE; i++)
+		rte_free(vhost_txbuff[i * MAX_VHOST_DEVICE + vid]);
+
 	if (builtin_net_driver)
 		vs_vhost_net_remove(vdev);
 

@@ -1298,6 +1390,7 @@ static int
 new_device(int vid)
 {
 	int lcore, core_add = 0;
+	uint16_t i;
 	uint32_t device_num_min = num_devices;
 	struct vhost_dev *vdev;
 	vdev = rte_zmalloc("vhost device", sizeof(*vdev), RTE_CACHE_LINE_SIZE);

@@ -1309,6 +1402,19 @@ new_device(int vid)
 	}
 	vdev->vid = vid;
 
+	for (i = 0; i < RTE_MAX_LCORE; i++) {
+		vhost_txbuff[i * MAX_VHOST_DEVICE + vid]
+			= rte_zmalloc("vhost bufftable",
+				sizeof(struct vhost_bufftable),
+				RTE_CACHE_LINE_SIZE);
+
+		if (vhost_txbuff[i * MAX_VHOST_DEVICE + vid] == NULL) {
+			RTE_LOG(INFO, VHOST_DATA,
+			  "(%d) couldn't allocate memory for vhost TX\n", vid);
+			return -1;
+		}
+	}
+
 	if (builtin_net_driver)
 		vs_vhost_net_setup(vdev);
 

@@ -1343,12 +1449,15 @@ new_device(int vid)
 	if (async_vhost_driver) {
 		struct rte_vhost_async_features f;
 		struct rte_vhost_async_channel_ops channel_ops;
+
 		if (strncmp(dma_type, "ioat", 4) == 0) {
 			channel_ops.transfer_data = ioat_transfer_data_cb;
 			channel_ops.check_completed_copies =
 				ioat_check_completed_copies_cb;
+
 			f.async_inorder = 1;
 			f.async_threshold = 256;
+
 			return rte_vhost_async_channel_register(vid, VIRTIO_RXQ,
 				f.intval, &channel_ops);
 		}

@@ -1392,8 +1501,8 @@ print_stats(__rte_unused void *arg)
 			tx         = vdev->stats.tx;
 			tx_dropped = tx_total - tx;
 
-			rx_total   = rte_atomic64_read(&vdev->stats.rx_total_atomic);
-			rx         = rte_atomic64_read(&vdev->stats.rx_atomic);
+			rx_total   = atomic_load(&vdev->stats.rx_total_atomic);
+			rx         = atomic_load(&vdev->stats.rx_atomic);
 			rx_dropped = rx_total - rx;
 
 			printf("Statistics for device %d\n"

@@ -1592,6 +1701,7 @@ main(int argc, char *argv[])
 	/* Register vhost user driver to handle vhost messages. */
 	for (i = 0; i < nb_sockets; i++) {
 		char *file = socket_files + i * PATH_MAX;
+
 		if (async_vhost_driver)
 			flags = flags | RTE_VHOST_USER_ASYNC_COPY;
 

index 4317b6ae8..6aa798a3e 100644
@@ -8,6 +8,7 @@
 #include <sys/queue.h>
 
 #include <rte_ether.h>
+#include <stdatomic.h>
 
 /* Macros for printing using RTE_LOG */
 #define RTE_LOGTYPE_VHOST_CONFIG RTE_LOGTYPE_USER1

@@ -21,8 +22,8 @@ enum {VIRTIO_RXQ, VIRTIO_TXQ, VIRTIO_QNUM};
 struct device_statistics {
 	uint64_t	tx;
 	uint64_t	tx_total;
-	rte_atomic64_t	rx_atomic;
-	rte_atomic64_t	rx_total_atomic;
+	atomic_int_least64_t	rx_atomic;
+	atomic_int_least64_t	rx_total_atomic;
 };
 
 struct vhost_queue {

@@ -51,7 +52,7 @@ struct vhost_dev {
 	uint64_t features;
 	size_t hdr_len;
 	uint16_t nr_vrings;
-	uint16_t nr_async_pkts;
+	atomic_int_least16_t nr_async_pkts;
 	struct rte_vhost_memory *mem;
 	struct device_statistics stats;
 	TAILQ_ENTRY(vhost_dev) global_vdev_entry;

index 71d8a1f1f5..679d1e2f58 100644
@@ -17,6 +17,7 @@ struct packet_tracker {
 	unsigned short next_read;
 	unsigned short next_write;
 	unsigned short last_remain;
+	unsigned short ioat_space;
 };
 
 struct packet_tracker cb_tracker[MAX_VHOST_DEVICE];

@@ -113,7 +114,7 @@ open_ioat(const char *value)
 			goto out;
 		}
 		rte_rawdev_start(dev_id);
-
+		cb_tracker[dev_id].ioat_space = IOAT_RING_SIZE;
 		dma_info->nr++;
 		i++;
 	}

@@ -140,13 +141,9 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 			src = descs[i_desc].src;
 			dst = descs[i_desc].dst;
 			i_seg = 0;
+			if (cb_tracker[dev_id].ioat_space < src->nr_segs)
+				break;
 			while (i_seg < src->nr_segs) {
-				/*
-				 * TODO: Assuming that the ring space of the
-				 * IOAT device is large enough, so there is no
-				 * error here, and the actual error handling
-				 * will be added later.
-				 */
 				rte_ioat_enqueue_copy(dev_id,
 					(uintptr_t)(src->iov[i_seg].iov_base)
 						+ src->offset,

@@ -158,7 +155,8 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 				i_seg++;
 			}
 			write &= mask;
-			cb_tracker[dev_id].size_track[write] = i_seg;
+			cb_tracker[dev_id].size_track[write] = src->nr_segs;
+			cb_tracker[dev_id].ioat_space -= src->nr_segs;
 			write++;
 		}
 	} else {

@@ -178,17 +176,21 @@ ioat_check_completed_copies_cb(int vid, uint16_t queue_id,
 {
 	if (!opaque_data) {
 		uintptr_t dump[255];
-		unsigned short n_seg;
+		int n_seg;
 		unsigned short read, write;
 		unsigned short nb_packet = 0;
 		unsigned short mask = MAX_ENQUEUED_SIZE - 1;
 		unsigned short i;
+
 		int dev_id = dma_bind[vid].dmas[queue_id * 2
 				+ VIRTIO_RXQ].dev_id;
 		n_seg = rte_ioat_completed_ops(dev_id, 255, dump, dump);
-		n_seg += cb_tracker[dev_id].last_remain;
 		if (!n_seg)
 			return 0;
+
+		cb_tracker[dev_id].ioat_space += n_seg;
+		n_seg += cb_tracker[dev_id].last_remain;
+
 		read = cb_tracker[dev_id].next_read;
 		write = cb_tracker[dev_id].next_write;
 		for (i = 0; i < max_packets; i++) {

index d6e1e2e079..0a1dbb8117 100644
@@ -11,7 +11,7 @@
 
 #define MAX_VHOST_DEVICE 1024
 #define IOAT_RING_SIZE 4096
-#define MAX_ENQUEUED_SIZE 256
+#define MAX_ENQUEUED_SIZE 512
 
 struct dma_info {
 	struct rte_pci_addr addr;

355 lines not on this page · raw message

index 4317b6ae81..6aa798a3e2 100644
@@ -8,6 +8,7 @@
 #include <sys/queue.h>
 
 #include <rte_ether.h>
+#include <stdatomic.h>
 
 /* Macros for printing using RTE_LOG */
 #define RTE_LOGTYPE_VHOST_CONFIG RTE_LOGTYPE_USER1

@@ -21,8 +22,8 @@ enum {VIRTIO_RXQ, VIRTIO_TXQ, VIRTIO_QNUM};
 struct device_statistics {
 	uint64_t	tx;
 	uint64_t	tx_total;
-	rte_atomic64_t	rx_atomic;
-	rte_atomic64_t	rx_total_atomic;
+	atomic_int_least64_t	rx_atomic;
+	atomic_int_least64_t	rx_total_atomic;
 };
 
 struct vhost_queue {

@@ -51,7 +52,7 @@ struct vhost_dev {
 	uint64_t features;
 	size_t hdr_len;
 	uint16_t nr_vrings;
-	uint16_t nr_async_pkts;
+	atomic_int_least16_t nr_async_pkts;
 	struct rte_vhost_memory *mem;
 	struct device_statistics stats;
 	TAILQ_ENTRY(vhost_dev) global_vdev_entry;

index 71d8a1f1f5..dbad28d43e 100644
@@ -17,6 +17,7 @@ struct packet_tracker {
 	unsigned short next_read;
 	unsigned short next_write;
 	unsigned short last_remain;
+	unsigned short ioat_space;
 };
 
 struct packet_tracker cb_tracker[MAX_VHOST_DEVICE];

@@ -113,7 +114,7 @@ open_ioat(const char *value)
 			goto out;
 		}
 		rte_rawdev_start(dev_id);
-
+		cb_tracker[dev_id].ioat_space = IOAT_RING_SIZE;
 		dma_info->nr++;
 		i++;
 	}

@@ -140,13 +141,9 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 			src = descs[i_desc].src;
 			dst = descs[i_desc].dst;
 			i_seg = 0;
+			if (cb_tracker[dev_id].ioat_space < src->nr_segs)
+				break;
 			while (i_seg < src->nr_segs) {
-				/*
-				 * TODO: Assuming that the ring space of the
-				 * IOAT device is large enough, so there is no
-				 * error here, and the actual error handling
-				 * will be added later.
-				 */
 				rte_ioat_enqueue_copy(dev_id,
 					(uintptr_t)(src->iov[i_seg].iov_base)
 						+ src->offset,

@@ -158,7 +155,8 @@ ioat_transfer_data_cb(int vid, uint16_t queue_id,
 				i_seg++;
 			}
 			write &= mask;
-			cb_tracker[dev_id].size_track[write] = i_seg;
+			cb_tracker[dev_id].size_track[write] = src->nr_segs;
+			cb_tracker[dev_id].ioat_space -= src->nr_segs;
 			write++;
 		}
 	} else {

@@ -178,17 +176,21 @@ ioat_check_completed_copies_cb(int vid, uint16_t queue_id,
 {
 	if (!opaque_data) {
 		uintptr_t dump[255];
-		unsigned short n_seg;
+		int n_seg;
 		unsigned short read, write;
 		unsigned short nb_packet = 0;
 		unsigned short mask = MAX_ENQUEUED_SIZE - 1;
 		unsigned short i;
+
 		int dev_id = dma_bind[vid].dmas[queue_id * 2
 				+ VIRTIO_RXQ].dev_id;
 		n_seg = rte_ioat_completed_ops(dev_id, 255, dump, dump);
-		n_seg += cb_tracker[dev_id].last_remain;
-		if (!n_seg)
+		if (n_seg <= 0)
 			return 0;
+
+		cb_tracker[dev_id].ioat_space += n_seg;
+		n_seg += cb_tracker[dev_id].last_remain;
+
 		read = cb_tracker[dev_id].next_read;
 		write = cb_tracker[dev_id].next_write;
 		for (i = 0; i < max_packets; i++) {