examples/vhost: refactor vhost data path
authorCheng Jiang <cheng1.jiang@intel.com>
Tue, 12 Jan 2021 04:38:57 +0000 (04:38 +0000)
committerFerruh Yigit <ferruh.yigit@intel.com>
Fri, 29 Jan 2021 17:16:09 +0000 (18:16 +0100)
Change the vm2vm data path to batch enqueue for better performance.
Support latest async vhost API, refactor vhost async data path,
replace rte_atomicNN_xxx to __atomic_XXX and clean some codes.

Signed-off-by: Cheng Jiang <cheng1.jiang@intel.com>
Reviewed-by: Jiayu Hu <jiayu.hu@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
examples/vhost/ioat.h
examples/vhost/main.c
examples/vhost/main.h

index d6e1e2e..0a1dbb8 100644 (file)
@@ -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;
index 2230997..233a2dc 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
@@ -804,43 +816,118 @@ 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
+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_sub_fetch(&vdev->nr_async_pkts, complete_count,
+                       __ATOMIC_SEQ_CST);
+               free_pkts(p_cpl, complete_count);
+       }
+}
+
 static __rte_always_inline void
-virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
+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], *comp_pkt;
-       uint32_t nr_comp = 0;
 
        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, &comp_pkt, &nr_comp);
-               if (nr_comp == 1)
-                       goto done;
-
-               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);
        }
 
-done:
        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;
+       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_add_fetch(&vdev->nr_async_pkts, ret - cpu_cpl_nr,
+                               __ATOMIC_SEQ_CST);
+
+               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.
@@ -850,7 +937,8 @@ 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);
@@ -873,7 +961,19 @@ 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;
 }
 
@@ -944,13 +1044,6 @@ 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)
 {
@@ -983,16 +1076,14 @@ 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,
@@ -1077,32 +1168,15 @@ 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;
-       struct rte_mbuf *pkts[MAX_PKT_BURST], *comp_pkts[MAX_PKT_BURST];
-       uint32_t nr_comp = 0;
+       struct rte_mbuf *pkts[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;
 
@@ -1128,22 +1202,34 @@ 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, comp_pkts,
-                                       &nr_comp);
-               if (nr_comp > 0) {
-                       free_pkts(comp_pkts, nr_comp);
-                       enqueue_count -= nr_comp;
-               }
-               vdev->nr_async_pkts += enqueue_count;
+                                       VIRTIO_RXQ, pkts, rx_count,
+                                       m_cpu_cpl, &cpu_cpl_nr);
+               __atomic_add_fetch(&vdev->nr_async_pkts,
+                                       enqueue_count - cpu_cpl_nr,
+                                       __ATOMIC_SEQ_CST);
+               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)
@@ -1212,7 +1298,7 @@ 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.
@@ -1253,6 +1339,7 @@ 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)
@@ -1266,6 +1353,9 @@ 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);
 
@@ -1308,6 +1398,7 @@ 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);
@@ -1319,6 +1410,19 @@ 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);
 
@@ -1353,12 +1457,15 @@ 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);
                }
@@ -1402,8 +1509,10 @@ 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"
@@ -1602,6 +1711,7 @@ 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 4317b6a..2d6c05f 100644 (file)
@@ -21,8 +21,8 @@ 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;
+       uint64_t        rx_atomic;
+       uint64_t        rx_total_atomic;
 };
 
 struct vhost_queue {
@@ -51,7 +51,7 @@ struct vhost_dev {
        uint64_t features;
        size_t hdr_len;
        uint16_t nr_vrings;
-       uint16_t nr_async_pkts;
+       uint64_t nr_async_pkts;
        struct rte_vhost_memory *mem;
        struct device_statistics stats;
        TAILQ_ENTRY(vhost_dev) global_vdev_entry;