examples/vhost: enhance getopt_long usage
[dpdk.git] / examples / vhost / main.c
index ec88874..2ca7d98 100644 (file)
@@ -14,7 +14,6 @@
 #include <sys/param.h>
 #include <unistd.h>
 
-#include <rte_atomic.h>
 #include <rte_cycles.h>
 #include <rte_ethdev.h>
 #include <rte_log.h>
@@ -179,9 +178,22 @@ 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
@@ -466,6 +478,35 @@ us_vhost_usage(const char *prgname)
               prgname);
 }
 
+enum {
+#define OPT_VM2VM               "vm2vm"
+       OPT_VM2VM_NUM = 256,
+#define OPT_RX_RETRY            "rx-retry"
+       OPT_RX_RETRY_NUM,
+#define OPT_RX_RETRY_DELAY      "rx-retry-delay"
+       OPT_RX_RETRY_DELAY_NUM,
+#define OPT_RX_RETRY_NUMB       "rx-retry-num"
+       OPT_RX_RETRY_NUMB_NUM,
+#define OPT_MERGEABLE           "mergeable"
+       OPT_MERGEABLE_NUM,
+#define OPT_STATS               "stats"
+       OPT_STATS_NUM,
+#define OPT_SOCKET_FILE         "socket-file"
+       OPT_SOCKET_FILE_NUM,
+#define OPT_TX_CSUM             "tx-csum"
+       OPT_TX_CSUM_NUM,
+#define OPT_TSO                 "tso"
+       OPT_TSO_NUM,
+#define OPT_CLIENT              "client"
+       OPT_CLIENT_NUM,
+#define OPT_BUILTIN_NET_DRIVER  "builtin-net-driver"
+       OPT_BUILTIN_NET_DRIVER_NUM,
+#define OPT_DMA_TYPE            "dma-type"
+       OPT_DMA_TYPE_NUM,
+#define OPT_DMAS                "dmas"
+       OPT_DMAS_NUM,
+};
+
 /*
  * Parse the arguments given in the command line of the application.
  */
@@ -477,19 +518,32 @@ us_vhost_parse_args(int argc, char **argv)
        unsigned i;
        const char *prgname = argv[0];
        static struct option long_option[] = {
-               {"vm2vm", required_argument, NULL, 0},
-               {"rx-retry", required_argument, NULL, 0},
-               {"rx-retry-delay", required_argument, NULL, 0},
-               {"rx-retry-num", required_argument, NULL, 0},
-               {"mergeable", required_argument, NULL, 0},
-               {"stats", required_argument, NULL, 0},
-               {"socket-file", required_argument, NULL, 0},
-               {"tx-csum", required_argument, NULL, 0},
-               {"tso", required_argument, NULL, 0},
-               {"client", no_argument, &client_mode, 1},
-               {"builtin-net-driver", no_argument, &builtin_net_driver, 1},
-               {"dma-type", required_argument, NULL, 0},
-               {"dmas", required_argument, NULL, 0},
+               {OPT_VM2VM, required_argument,
+                               NULL, OPT_VM2VM_NUM},
+               {OPT_RX_RETRY, required_argument,
+                               NULL, OPT_RX_RETRY_NUM},
+               {OPT_RX_RETRY_DELAY, required_argument,
+                               NULL, OPT_RX_RETRY_DELAY_NUM},
+               {OPT_RX_RETRY_NUMB, required_argument,
+                               NULL, OPT_RX_RETRY_NUMB_NUM},
+               {OPT_MERGEABLE, required_argument,
+                               NULL, OPT_MERGEABLE_NUM},
+               {OPT_STATS, required_argument,
+                               NULL, OPT_STATS_NUM},
+               {OPT_SOCKET_FILE, required_argument,
+                               NULL, OPT_SOCKET_FILE_NUM},
+               {OPT_TX_CSUM, required_argument,
+                               NULL, OPT_TX_CSUM_NUM},
+               {OPT_TSO, required_argument,
+                               NULL, OPT_TSO_NUM},
+               {OPT_CLIENT, no_argument,
+                               NULL, OPT_CLIENT_NUM},
+               {OPT_BUILTIN_NET_DRIVER, no_argument,
+                               NULL, OPT_BUILTIN_NET_DRIVER_NUM},
+               {OPT_DMA_TYPE, required_argument,
+                               NULL, OPT_DMA_TYPE_NUM},
+               {OPT_DMAS, required_argument,
+                               NULL, OPT_DMAS_NUM},
                {NULL, 0, 0, 0},
        };
 
@@ -512,151 +566,131 @@ us_vhost_parse_args(int argc, char **argv)
                        vmdq_conf_default.rx_adv_conf.vmdq_rx_conf.rx_mode =
                                ETH_VMDQ_ACCEPT_BROADCAST |
                                ETH_VMDQ_ACCEPT_MULTICAST;
-
                        break;
 
-               case 0:
-                       /* Enable/disable vm2vm comms. */
-                       if (!strncmp(long_option[option_index].name, "vm2vm",
-                               MAX_LONG_OPT_SZ)) {
-                               ret = parse_num_opt(optarg, (VM2VM_LAST - 1));
-                               if (ret == -1) {
-                                       RTE_LOG(INFO, VHOST_CONFIG,
-                                               "Invalid argument for "
-                                               "vm2vm [0|1|2]\n");
-                                       us_vhost_usage(prgname);
-                                       return -1;
-                               } else {
-                                       vm2vm_mode = (vm2vm_type)ret;
-                               }
+               case OPT_VM2VM_NUM:
+                       ret = parse_num_opt(optarg, (VM2VM_LAST - 1));
+                       if (ret == -1) {
+                               RTE_LOG(INFO, VHOST_CONFIG,
+                                       "Invalid argument for "
+                                       "vm2vm [0|1|2]\n");
+                               us_vhost_usage(prgname);
+                               return -1;
                        }
+                       vm2vm_mode = (vm2vm_type)ret;
+                       break;
 
-                       /* Enable/disable retries on RX. */
-                       if (!strncmp(long_option[option_index].name, "rx-retry", MAX_LONG_OPT_SZ)) {
-                               ret = parse_num_opt(optarg, 1);
-                               if (ret == -1) {
-                                       RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for rx-retry [0|1]\n");
-                                       us_vhost_usage(prgname);
-                                       return -1;
-                               } else {
-                                       enable_retry = ret;
-                               }
+               case OPT_RX_RETRY_NUM:
+                       ret = parse_num_opt(optarg, 1);
+                       if (ret == -1) {
+                               RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for rx-retry [0|1]\n");
+                               us_vhost_usage(prgname);
+                               return -1;
                        }
+                       enable_retry = ret;
+                       break;
 
-                       /* Enable/disable TX checksum offload. */
-                       if (!strncmp(long_option[option_index].name, "tx-csum", MAX_LONG_OPT_SZ)) {
-                               ret = parse_num_opt(optarg, 1);
-                               if (ret == -1) {
-                                       RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for tx-csum [0|1]\n");
-                                       us_vhost_usage(prgname);
-                                       return -1;
-                               } else
-                                       enable_tx_csum = ret;
+               case OPT_TX_CSUM_NUM:
+                       ret = parse_num_opt(optarg, 1);
+                       if (ret == -1) {
+                               RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for tx-csum [0|1]\n");
+                               us_vhost_usage(prgname);
+                               return -1;
                        }
+                       enable_tx_csum = ret;
+                       break;
 
-                       /* Enable/disable TSO offload. */
-                       if (!strncmp(long_option[option_index].name, "tso", MAX_LONG_OPT_SZ)) {
-                               ret = parse_num_opt(optarg, 1);
-                               if (ret == -1) {
-                                       RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for tso [0|1]\n");
-                                       us_vhost_usage(prgname);
-                                       return -1;
-                               } else
-                                       enable_tso = ret;
+               case OPT_TSO_NUM:
+                       ret = parse_num_opt(optarg, 1);
+                       if (ret == -1) {
+                               RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for tso [0|1]\n");
+                               us_vhost_usage(prgname);
+                               return -1;
                        }
+                       enable_tso = ret;
+                       break;
 
-                       /* Specify the retries delay time (in useconds) on RX. */
-                       if (!strncmp(long_option[option_index].name, "rx-retry-delay", MAX_LONG_OPT_SZ)) {
-                               ret = parse_num_opt(optarg, INT32_MAX);
-                               if (ret == -1) {
-                                       RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for rx-retry-delay [0-N]\n");
-                                       us_vhost_usage(prgname);
-                                       return -1;
-                               } else {
-                                       burst_rx_delay_time = ret;
-                               }
+               case OPT_RX_RETRY_DELAY_NUM:
+                       ret = parse_num_opt(optarg, INT32_MAX);
+                       if (ret == -1) {
+                               RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for rx-retry-delay [0-N]\n");
+                               us_vhost_usage(prgname);
+                               return -1;
                        }
+                       burst_rx_delay_time = ret;
+                       break;
 
-                       /* Specify the retries number on RX. */
-                       if (!strncmp(long_option[option_index].name, "rx-retry-num", MAX_LONG_OPT_SZ)) {
-                               ret = parse_num_opt(optarg, INT32_MAX);
-                               if (ret == -1) {
-                                       RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for rx-retry-num [0-N]\n");
-                                       us_vhost_usage(prgname);
-                                       return -1;
-                               } else {
-                                       burst_rx_retry_num = ret;
-                               }
+               case OPT_RX_RETRY_NUMB_NUM:
+                       ret = parse_num_opt(optarg, INT32_MAX);
+                       if (ret == -1) {
+                               RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for rx-retry-num [0-N]\n");
+                               us_vhost_usage(prgname);
+                               return -1;
                        }
+                       burst_rx_retry_num = ret;
+                       break;
 
-                       /* Enable/disable RX mergeable buffers. */
-                       if (!strncmp(long_option[option_index].name, "mergeable", MAX_LONG_OPT_SZ)) {
-                               ret = parse_num_opt(optarg, 1);
-                               if (ret == -1) {
-                                       RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for mergeable [0|1]\n");
-                                       us_vhost_usage(prgname);
-                                       return -1;
-                               } else {
-                                       mergeable = !!ret;
-                                       if (ret) {
-                                               vmdq_conf_default.rxmode.offloads |=
-                                                       DEV_RX_OFFLOAD_JUMBO_FRAME;
-                                               vmdq_conf_default.rxmode.max_rx_pkt_len
-                                                       = JUMBO_FRAME_MAX_SIZE;
-                                       }
-                               }
+               case OPT_MERGEABLE_NUM:
+                       ret = parse_num_opt(optarg, 1);
+                       if (ret == -1) {
+                               RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for mergeable [0|1]\n");
+                               us_vhost_usage(prgname);
+                               return -1;
                        }
-
-                       /* Enable/disable stats. */
-                       if (!strncmp(long_option[option_index].name, "stats", MAX_LONG_OPT_SZ)) {
-                               ret = parse_num_opt(optarg, INT32_MAX);
-                               if (ret == -1) {
-                                       RTE_LOG(INFO, VHOST_CONFIG,
-                                               "Invalid argument for stats [0..N]\n");
-                                       us_vhost_usage(prgname);
-                                       return -1;
-                               } else {
-                                       enable_stats = ret;
-                               }
+                       mergeable = !!ret;
+                       if (ret) {
+                               vmdq_conf_default.rxmode.offloads |=
+                                       DEV_RX_OFFLOAD_JUMBO_FRAME;
+                               vmdq_conf_default.rxmode.max_rx_pkt_len
+                                       = JUMBO_FRAME_MAX_SIZE;
                        }
+                       break;
 
-                       /* Set socket file path. */
-                       if (!strncmp(long_option[option_index].name,
-                                               "socket-file", MAX_LONG_OPT_SZ)) {
-                               if (us_vhost_parse_socket_path(optarg) == -1) {
-                                       RTE_LOG(INFO, VHOST_CONFIG,
-                                       "Invalid argument for socket name (Max %d characters)\n",
-                                       PATH_MAX);
-                                       us_vhost_usage(prgname);
-                                       return -1;
-                               }
+               case OPT_STATS_NUM:
+                       ret = parse_num_opt(optarg, INT32_MAX);
+                       if (ret == -1) {
+                               RTE_LOG(INFO, VHOST_CONFIG,
+                                       "Invalid argument for stats [0..N]\n");
+                               us_vhost_usage(prgname);
+                               return -1;
                        }
+                       enable_stats = ret;
+                       break;
 
-                       if (!strncmp(long_option[option_index].name,
-                                               "dma-type", MAX_LONG_OPT_SZ)) {
-                               if (strlen(optarg) >= MAX_LONG_OPT_SZ) {
-                                       RTE_LOG(INFO, VHOST_CONFIG,
-                                               "Wrong DMA type\n");
-                                       us_vhost_usage(prgname);
-                                       return -1;
-                               }
-                               strcpy(dma_type, optarg);
+               /* Set socket file path. */
+               case OPT_SOCKET_FILE_NUM:
+                       if (us_vhost_parse_socket_path(optarg) == -1) {
+                               RTE_LOG(INFO, VHOST_CONFIG,
+                               "Invalid argument for socket name (Max %d characters)\n",
+                               PATH_MAX);
+                               us_vhost_usage(prgname);
+                               return -1;
                        }
+                       break;
 
-                       if (!strncmp(long_option[option_index].name,
-                                               "dmas", MAX_LONG_OPT_SZ)) {
-                               if (open_dma(optarg) == -1) {
-                                       RTE_LOG(INFO, VHOST_CONFIG,
-                                               "Wrong DMA args\n");
-                                       us_vhost_usage(prgname);
-                                       return -1;
-                               }
-                               async_vhost_driver = 1;
+               case OPT_DMA_TYPE_NUM:
+                       strcpy(dma_type, optarg);
+                       break;
+
+               case OPT_DMAS_NUM:
+                       if (open_dma(optarg) == -1) {
+                               RTE_LOG(INFO, VHOST_CONFIG,
+                                       "Wrong DMA args\n");
+                               us_vhost_usage(prgname);
+                               return -1;
                        }
+                       async_vhost_driver = 1;
+                       break;
+
+               case OPT_CLIENT_NUM:
+                       client_mode = 1;
+                       break;
 
+               case OPT_BUILTIN_NET_DRIVER_NUM:
+                       builtin_net_driver = 1;
                        break;
 
-                       /* Invalid option - print options. */
+               /* Invalid option - print options. */
                default:
                        us_vhost_usage(prgname);
                        return -1;
@@ -804,39 +838,113 @@ 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)
+               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_add_fetch(&dst_vdev->stats.rx_total_atomic, 1,
+                               __ATOMIC_SEQ_CST);
+               __atomic_add_fetch(&dst_vdev->stats.rx_atomic, ret,
+                               __ATOMIC_SEQ_CST);
                src_vdev->stats.tx_total++;
                src_vdev->stats.tx += ret;
        }
 }
 
+static __rte_always_inline void
+drain_vhost(struct vhost_dev *vdev)
+{
+       uint16_t ret;
+       uint32_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);
+
+               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_add_fetch(&vdev->stats.rx_total_atomic, nr_xmit,
+                               __ATOMIC_SEQ_CST);
+               __atomic_add_fetch(&vdev->stats.rx_atomic, ret,
+                               __ATOMIC_SEQ_CST);
+       }
+
+       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 +954,8 @@ 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 +978,19 @@ 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 +1061,6 @@ 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 +1093,14 @@ 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 +1185,6 @@ 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 +1194,6 @@ 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 +1219,31 @@ 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);
+               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_add_fetch(&vdev->stats.rx_total_atomic, rx_count,
+                               __ATOMIC_SEQ_CST);
+               __atomic_add_fetch(&vdev->stats.rx_atomic, enqueue_count,
+                               __ATOMIC_SEQ_CST);
        }
 
        if (!async_vhost_driver)
@@ -1202,7 +1312,7 @@ 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 +1353,7 @@ 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 +1367,9 @@ 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);
 
@@ -1298,15 +1412,9 @@ 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;
-
-       struct rte_vhost_async_channel_ops channel_ops = {
-               .transfer_data = ioat_transfer_data_cb,
-               .check_completed_copies = ioat_check_completed_copies_cb
-       };
-       struct rte_vhost_async_features f;
-
        vdev = rte_zmalloc("vhost device", sizeof(*vdev), RTE_CACHE_LINE_SIZE);
        if (vdev == NULL) {
                RTE_LOG(INFO, VHOST_DATA,
@@ -1316,6 +1424,19 @@ 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);
 
@@ -1348,10 +1469,20 @@ new_device(int vid)
                vid, vdev->coreid);
 
        if (async_vhost_driver) {
-               f.async_inorder = 1;
-               f.async_threshold = 256;
-               return rte_vhost_async_channel_register(vid, VIRTIO_RXQ,
-                       f.intval, &channel_ops);
+               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);
+               }
        }
 
        return 0;
@@ -1392,8 +1523,10 @@ 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_n(&vdev->stats.rx_total_atomic,
+                               __ATOMIC_SEQ_CST);
+                       rx         = __atomic_load_n(&vdev->stats.rx_atomic,
+                               __ATOMIC_SEQ_CST);
                        rx_dropped = rx_total - rx;
 
                        printf("Statistics for device %d\n"
@@ -1592,6 +1725,7 @@ 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;