vhost: access header only if offloading is supported
authorMaxime Coquelin <maxime.coquelin@redhat.com>
Fri, 14 Oct 2016 08:07:07 +0000 (10:07 +0200)
committerYuanhan Liu <yuanhan.liu@linux.intel.com>
Wed, 26 Oct 2016 11:39:09 +0000 (13:39 +0200)
If offloading features are not negotiated, parsing the virtio header
is not needed.

Micro-benchmark with testpmd shows that the gain is +4% with indirect
descriptors, +1% when using direct descriptors.

Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
lib/librte_vhost/virtio_net.c

index 812e5d3..15ef0b0 100644 (file)
@@ -555,6 +555,18 @@ rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
                return virtio_dev_rx(dev, queue_id, pkts, count);
 }
 
+static inline bool
+virtio_net_with_host_offload(struct virtio_net *dev)
+{
+       if (dev->features &
+                       (VIRTIO_NET_F_CSUM | VIRTIO_NET_F_HOST_ECN |
+                        VIRTIO_NET_F_HOST_TSO4 | VIRTIO_NET_F_HOST_TSO6 |
+                        VIRTIO_NET_F_HOST_UFO))
+               return true;
+
+       return false;
+}
+
 static void
 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 {
@@ -607,6 +619,9 @@ vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
        void *l4_hdr = NULL;
        struct tcp_hdr *tcp_hdr = NULL;
 
+       if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
+               return;
+
        parse_ethernet(m, &l4_proto, &l4_hdr);
        if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
                if (hdr->csum_start == (m->l2_len + m->l3_len)) {
@@ -702,7 +717,7 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
        uint32_t mbuf_avail, mbuf_offset;
        uint32_t cpy_len;
        struct rte_mbuf *cur = m, *prev = m;
-       struct virtio_net_hdr *hdr;
+       struct virtio_net_hdr *hdr = NULL;
        /* A counter to avoid desc dead loop chain */
        uint32_t nr_desc = 1;
 
@@ -715,8 +730,10 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
        if (unlikely(!desc_addr))
                return -1;
 
-       hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
-       rte_prefetch0(hdr);
+       if (virtio_net_with_host_offload(dev)) {
+               hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
+               rte_prefetch0(hdr);
+       }
 
        /*
         * A virtio driver normally uses at least 2 desc buffers
@@ -733,18 +750,18 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
                if (unlikely(!desc_addr))
                        return -1;
 
-               rte_prefetch0((void *)(uintptr_t)desc_addr);
-
                desc_offset = 0;
                desc_avail  = desc->len;
                nr_desc    += 1;
-
-               PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
        } else {
                desc_avail  = desc->len - dev->vhost_hlen;
                desc_offset = dev->vhost_hlen;
        }
 
+       rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
+
+       PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
+
        mbuf_offset = 0;
        mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
        while (1) {
@@ -831,7 +848,7 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
        prev->data_len = mbuf_offset;
        m->pkt_len    += mbuf_offset;
 
-       if (hdr->flags != 0 || hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE)
+       if (hdr)
                vhost_dequeue_offload(hdr, m);
 
        return 0;