vhost: add inflight description
[dpdk.git] / lib / librte_vhost / virtio_net.c
index 40a2923..beba9b9 100644 (file)
@@ -110,8 +110,6 @@ flush_shadow_used_ring_packed(struct virtio_net *dev,
                        used_idx -= vq->size;
        }
 
-       rte_smp_wmb();
-
        for (i = 0; i < vq->shadow_used_idx; i++) {
                uint16_t flags;
 
@@ -147,7 +145,8 @@ flush_shadow_used_ring_packed(struct virtio_net *dev,
                }
        }
 
-       vq->desc_packed[head_idx].flags = head_flags;
+       __atomic_store_n(&vq->desc_packed[head_idx].flags, head_flags,
+                        __ATOMIC_RELEASE);
 
        vhost_log_cache_used_vring(dev, vq,
                                head_idx *
@@ -178,7 +177,8 @@ do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 
        for (i = 0; i < count; i++) {
                rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
-               vhost_log_cache_write(dev, vq, elem[i].log_addr, elem[i].len);
+               vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
+                                          elem[i].len);
                PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
        }
 
@@ -218,15 +218,15 @@ virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 
                switch (csum_l4) {
                case PKT_TX_TCP_CKSUM:
-                       net_hdr->csum_offset = (offsetof(struct tcp_hdr,
+                       net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
                                                cksum));
                        break;
                case PKT_TX_UDP_CKSUM:
-                       net_hdr->csum_offset = (offsetof(struct udp_hdr,
+                       net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
                                                dgram_cksum));
                        break;
                case PKT_TX_SCTP_CKSUM:
-                       net_hdr->csum_offset = (offsetof(struct sctp_hdr,
+                       net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
                                                cksum));
                        break;
                }
@@ -238,9 +238,9 @@ virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 
        /* IP cksum verification cannot be bypassed, then calculate here */
        if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
-               struct ipv4_hdr *ipv4_hdr;
+               struct rte_ipv4_hdr *ipv4_hdr;
 
-               ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
+               ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
                                                   m_buf->l2_len);
                ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
        }
@@ -286,6 +286,8 @@ map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
                if (unlikely(!desc_addr))
                        return -1;
 
+               rte_prefetch0((void *)(uintptr_t)desc_addr);
+
                buf_vec[vec_id].buf_iova = desc_iova;
                buf_vec[vec_id].buf_addr = desc_addr;
                buf_vec[vec_id].buf_len  = desc_chunck_len;
@@ -337,7 +339,7 @@ fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
                         * The indirect desc table is not contiguous
                         * in process VA space, we have to copy it.
                         */
-                       idesc = alloc_copy_ind_table(dev, vq,
+                       idesc = vhost_alloc_copy_ind_table(dev, vq,
                                        vq->desc[idx].addr, vq->desc[idx].len);
                        if (unlikely(!idesc))
                                return -1;
@@ -454,7 +456,8 @@ fill_vec_buf_packed_indirect(struct virtio_net *dev,
                 * The indirect desc table is not contiguous
                 * in process VA space, we have to copy it.
                 */
-               idescs = alloc_copy_ind_table(dev, vq, desc->addr, desc->len);
+               idescs = vhost_alloc_copy_ind_table(dev,
+                               vq, desc->addr, desc->len);
                if (unlikely(!idescs))
                        return -1;
 
@@ -500,14 +503,13 @@ fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
        if (avail_idx < vq->last_avail_idx)
                wrap_counter ^= 1;
 
-       if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
-               return -1;
-
        /*
-        * The ordering between desc flags and desc
-        * content reads need to be enforced.
+        * Perform a load-acquire barrier in desc_is_avail to
+        * enforce the ordering between desc flags and desc
+        * content.
         */
-       rte_smp_rmb();
+       if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
+               return -1;
 
        *desc_count = 0;
        *len = 0;
@@ -610,6 +612,36 @@ reserve_avail_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
        return 0;
 }
 
+static __rte_noinline void
+copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
+               struct buf_vector *buf_vec,
+               struct virtio_net_hdr_mrg_rxbuf *hdr)
+{
+       uint64_t len;
+       uint64_t remain = dev->vhost_hlen;
+       uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
+       uint64_t iova = buf_vec->buf_iova;
+
+       while (remain) {
+               len = RTE_MIN(remain,
+                               buf_vec->buf_len);
+               dst = buf_vec->buf_addr;
+               rte_memcpy((void *)(uintptr_t)dst,
+                               (void *)(uintptr_t)src,
+                               len);
+
+               PRINT_PACKET(dev, (uintptr_t)dst,
+                               (uint32_t)len, 0);
+               vhost_log_cache_write_iova(dev, vq,
+                               iova, len);
+
+               remain -= len;
+               iova += len;
+               src += len;
+               buf_vec++;
+       }
+}
+
 static __rte_always_inline int
 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
                            struct rte_mbuf *m, struct buf_vector *buf_vec,
@@ -635,9 +667,6 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
        buf_iova = buf_vec[vec_idx].buf_iova;
        buf_len = buf_vec[vec_idx].buf_len;
 
-       if (nr_vec > 1)
-               rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
-
        if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
                error = -1;
                goto out;
@@ -680,10 +709,6 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
                        buf_iova = buf_vec[vec_idx].buf_iova;
                        buf_len = buf_vec[vec_idx].buf_len;
 
-                       /* Prefetch next buffer address. */
-                       if (vec_idx + 1 < nr_vec)
-                               rte_prefetch0((void *)(uintptr_t)
-                                               buf_vec[vec_idx + 1].buf_addr);
                        buf_offset = 0;
                        buf_avail  = buf_len;
                }
@@ -703,34 +728,11 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
                                                num_buffers);
 
                        if (unlikely(hdr == &tmp_hdr)) {
-                               uint64_t len;
-                               uint64_t remain = dev->vhost_hlen;
-                               uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
-                               uint64_t iova = buf_vec[0].buf_iova;
-                               uint16_t hdr_vec_idx = 0;
-
-                               while (remain) {
-                                       len = RTE_MIN(remain,
-                                               buf_vec[hdr_vec_idx].buf_len);
-                                       dst = buf_vec[hdr_vec_idx].buf_addr;
-                                       rte_memcpy((void *)(uintptr_t)dst,
-                                                       (void *)(uintptr_t)src,
-                                                       len);
-
-                                       PRINT_PACKET(dev, (uintptr_t)dst,
-                                                       (uint32_t)len, 0);
-                                       vhost_log_cache_write(dev, vq,
-                                                       iova, len);
-
-                                       remain -= len;
-                                       iova += len;
-                                       src += len;
-                                       hdr_vec_idx++;
-                               }
+                               copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
                        } else {
                                PRINT_PACKET(dev, (uintptr_t)hdr_addr,
                                                dev->vhost_hlen, 0);
-                               vhost_log_cache_write(dev, vq,
+                               vhost_log_cache_write_iova(dev, vq,
                                                buf_vec[0].buf_iova,
                                                dev->vhost_hlen);
                        }
@@ -745,8 +747,9 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
                        rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
                                rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
                                cpy_len);
-                       vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
-                                       cpy_len);
+                       vhost_log_cache_write_iova(dev, vq,
+                                                  buf_iova + buf_offset,
+                                                  cpy_len);
                        PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
                                cpy_len, 0);
                } else {
@@ -771,7 +774,7 @@ out:
        return error;
 }
 
-static __rte_always_inline uint32_t
+static __rte_noinline uint32_t
 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
        struct rte_mbuf **pkts, uint32_t count)
 {
@@ -804,8 +807,6 @@ virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
                        break;
                }
 
-               rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
-
                VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
                        dev->vid, vq->last_avail_idx,
                        vq->last_avail_idx + num_buffers);
@@ -830,7 +831,7 @@ virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
        return pkt_idx;
 }
 
-static __rte_always_inline uint32_t
+static __rte_noinline uint32_t
 virtio_dev_rx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
        struct rte_mbuf **pkts, uint32_t count)
 {
@@ -853,8 +854,6 @@ virtio_dev_rx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
                        break;
                }
 
-               rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
-
                VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
                        dev->vid, vq->last_avail_idx,
                        vq->last_avail_idx + num_buffers);
@@ -966,38 +965,39 @@ virtio_net_with_host_offload(struct virtio_net *dev)
 static void
 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 {
-       struct ipv4_hdr *ipv4_hdr;
-       struct ipv6_hdr *ipv6_hdr;
+       struct rte_ipv4_hdr *ipv4_hdr;
+       struct rte_ipv6_hdr *ipv6_hdr;
        void *l3_hdr = NULL;
-       struct ether_hdr *eth_hdr;
+       struct rte_ether_hdr *eth_hdr;
        uint16_t ethertype;
 
-       eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
+       eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
 
-       m->l2_len = sizeof(struct ether_hdr);
+       m->l2_len = sizeof(struct rte_ether_hdr);
        ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 
-       if (ethertype == ETHER_TYPE_VLAN) {
-               struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
+       if (ethertype == RTE_ETHER_TYPE_VLAN) {
+               struct rte_vlan_hdr *vlan_hdr =
+                       (struct rte_vlan_hdr *)(eth_hdr + 1);
 
-               m->l2_len += sizeof(struct vlan_hdr);
+               m->l2_len += sizeof(struct rte_vlan_hdr);
                ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
        }
 
        l3_hdr = (char *)eth_hdr + m->l2_len;
 
        switch (ethertype) {
-       case ETHER_TYPE_IPv4:
+       case RTE_ETHER_TYPE_IPV4:
                ipv4_hdr = l3_hdr;
                *l4_proto = ipv4_hdr->next_proto_id;
                m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
                *l4_hdr = (char *)l3_hdr + m->l3_len;
                m->ol_flags |= PKT_TX_IPV4;
                break;
-       case ETHER_TYPE_IPv6:
+       case RTE_ETHER_TYPE_IPV6:
                ipv6_hdr = l3_hdr;
                *l4_proto = ipv6_hdr->proto;
-               m->l3_len = sizeof(struct ipv6_hdr);
+               m->l3_len = sizeof(struct rte_ipv6_hdr);
                *l4_hdr = (char *)l3_hdr + m->l3_len;
                m->ol_flags |= PKT_TX_IPV6;
                break;
@@ -1014,7 +1014,7 @@ vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 {
        uint16_t l4_proto = 0;
        void *l4_hdr = NULL;
-       struct tcp_hdr *tcp_hdr = NULL;
+       struct rte_tcp_hdr *tcp_hdr = NULL;
 
        if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
                return;
@@ -1023,15 +1023,15 @@ vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
        if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
                if (hdr->csum_start == (m->l2_len + m->l3_len)) {
                        switch (hdr->csum_offset) {
-                       case (offsetof(struct tcp_hdr, cksum)):
+                       case (offsetof(struct rte_tcp_hdr, cksum)):
                                if (l4_proto == IPPROTO_TCP)
                                        m->ol_flags |= PKT_TX_TCP_CKSUM;
                                break;
-                       case (offsetof(struct udp_hdr, dgram_cksum)):
+                       case (offsetof(struct rte_udp_hdr, dgram_cksum)):
                                if (l4_proto == IPPROTO_UDP)
                                        m->ol_flags |= PKT_TX_UDP_CKSUM;
                                break;
-                       case (offsetof(struct sctp_hdr, cksum)):
+                       case (offsetof(struct rte_sctp_hdr, cksum)):
                                if (l4_proto == IPPROTO_SCTP)
                                        m->ol_flags |= PKT_TX_SCTP_CKSUM;
                                break;
@@ -1053,7 +1053,7 @@ vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
                case VIRTIO_NET_HDR_GSO_UDP:
                        m->ol_flags |= PKT_TX_UDP_SEG;
                        m->tso_segsz = hdr->gso_size;
-                       m->l4_len = sizeof(struct udp_hdr);
+                       m->l4_len = sizeof(struct rte_udp_hdr);
                        break;
                default:
                        RTE_LOG(WARNING, VHOST_DATA,
@@ -1063,10 +1063,25 @@ vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
        }
 }
 
-static __rte_always_inline void
-put_zmbuf(struct zcopy_mbuf *zmbuf)
+static __rte_noinline void
+copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
+               struct buf_vector *buf_vec)
 {
-       zmbuf->in_use = 0;
+       uint64_t len;
+       uint64_t remain = sizeof(struct virtio_net_hdr);
+       uint64_t src;
+       uint64_t dst = (uint64_t)(uintptr_t)hdr;
+
+       while (remain) {
+               len = RTE_MIN(remain, buf_vec->buf_len);
+               src = buf_vec->buf_addr;
+               rte_memcpy((void *)(uintptr_t)dst,
+                               (void *)(uintptr_t)src, len);
+
+               remain -= len;
+               dst += len;
+               buf_vec++;
+       }
 }
 
 static __rte_always_inline int
@@ -1095,37 +1110,16 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
                goto out;
        }
 
-       if (likely(nr_vec > 1))
-               rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
-
        if (virtio_net_with_host_offload(dev)) {
                if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
-                       uint64_t len;
-                       uint64_t remain = sizeof(struct virtio_net_hdr);
-                       uint64_t src;
-                       uint64_t dst = (uint64_t)(uintptr_t)&tmp_hdr;
-                       uint16_t hdr_vec_idx = 0;
-
                        /*
                         * No luck, the virtio-net header doesn't fit
                         * in a contiguous virtual area.
                         */
-                       while (remain) {
-                               len = RTE_MIN(remain,
-                                       buf_vec[hdr_vec_idx].buf_len);
-                               src = buf_vec[hdr_vec_idx].buf_addr;
-                               rte_memcpy((void *)(uintptr_t)dst,
-                                                  (void *)(uintptr_t)src, len);
-
-                               remain -= len;
-                               dst += len;
-                               hdr_vec_idx++;
-                       }
-
+                       copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
                        hdr = &tmp_hdr;
                } else {
                        hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
-                       rte_prefetch0(hdr);
                }
        }
 
@@ -1155,9 +1149,6 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
                buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
        }
 
-       rte_prefetch0((void *)(uintptr_t)
-                       (buf_addr + buf_offset));
-
        PRINT_PACKET(dev,
                        (uintptr_t)(buf_addr + buf_offset),
                        (uint32_t)buf_avail, 0);
@@ -1223,14 +1214,6 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
                        buf_iova = buf_vec[vec_idx].buf_iova;
                        buf_len = buf_vec[vec_idx].buf_len;
 
-                       /*
-                        * Prefecth desc n + 1 buffer while
-                        * desc n buffer is processed.
-                        */
-                       if (vec_idx + 1 < nr_vec)
-                               rte_prefetch0((void *)(uintptr_t)
-                                               buf_vec[vec_idx + 1].buf_addr);
-
                        buf_offset = 0;
                        buf_avail  = buf_len;
 
@@ -1306,7 +1289,7 @@ again:
        return NULL;
 }
 
-static __rte_always_inline uint16_t
+static __rte_noinline uint16_t
 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
        struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 {
@@ -1374,8 +1357,6 @@ virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
                if (likely(dev->dequeue_zero_copy == 0))
                        update_shadow_used_ring_split(vq, head_idx, 0);
 
-               rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
-
                pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
                if (unlikely(pkts[i] == NULL)) {
                        RTE_LOG(ERR, VHOST_DATA,
@@ -1428,7 +1409,7 @@ virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
        return i;
 }
 
-static __rte_always_inline uint16_t
+static __rte_noinline uint16_t
 virtio_dev_tx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
        struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 {
@@ -1485,8 +1466,6 @@ virtio_dev_tx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
                        update_shadow_used_ring_packed(vq, buf_id, 0,
                                        desc_count);
 
-               rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
-
                pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
                if (unlikely(pkts[i] == NULL)) {
                        RTE_LOG(ERR, VHOST_DATA,