On 06/30/18 19:06, Boris Pismenny wrote:


On 06/30/18 01:19, Willem de Bruijn wrote:
On Fri, Jun 29, 2018 at 2:24 AM Saeed Mahameed <sae...@mellanox.com> wrote:
From: Boris Pismenny <bor...@mellanox.com>

This patch enables UDP GSO support. We enable this by using two WQEs
the first is a UDP LSO WQE for all segments with equal length, and the
second is for the last segment in case it has different length.
Due to HW limitation, before sending, we must adjust the packet length fields.

We measure performance between two Intel(R) Xeon(R) CPU E5-2643 v2 @3.50GHz
machines connected back-to-back with Connectx4-Lx (40Gbps) NICs.
We compare single stream UDP, UDP GSO and UDP GSO with offload.
Performance:
                 | MSS (bytes)   | Throughput (Gbps)     | CPU utilization (%)
UDP GSO offload | 1472          | 35.6                  | 8%
UDP GSO         | 1472          | 25.5                  | 17%
UDP             | 1472          | 10.2                  | 17%
UDP GSO offload | 1024          | 35.6                  | 8%
UDP GSO         | 1024          | 19.2                  | 17%
UDP             | 1024          | 5.7                   | 17%
UDP GSO offload | 512           | 33.8                  | 16%
UDP GSO         | 512           | 10.4                  | 17%
UDP             | 512           | 3.5                   | 17%
Very nice results :)
Thanks. We expect to have 100Gbps results by netdev0x12.

+static void mlx5e_udp_gso_prepare_last_skb(struct sk_buff *skb,
+                                          struct sk_buff *nskb,
+                                          int remaining)
+{
+       int bytes_needed = remaining, remaining_headlen, remaining_page_offset; +       int headlen = skb_transport_offset(skb) + sizeof(struct udphdr);
+       int payload_len = remaining + sizeof(struct udphdr);
+       int k = 0, i, j;
+
+       skb_copy_bits(skb, 0, nskb->data, headlen);
+       nskb->dev = skb->dev;
+       skb_reset_mac_header(nskb);
+       skb_set_network_header(nskb, skb_network_offset(skb));
+       skb_set_transport_header(nskb, skb_transport_offset(skb));
+       skb_set_tail_pointer(nskb, headlen);
+
+       /* How many frags do we need? */
+       for (i = skb_shinfo(skb)->nr_frags - 1; i >= 0; i--) {
+               bytes_needed -= skb_frag_size(&skb_shinfo(skb)->frags[i]);
+               k++;
+               if (bytes_needed <= 0)
+                       break;
+       }
+
+       /* Fill the first frag and split it if necessary */
+       j = skb_shinfo(skb)->nr_frags - k;
+       remaining_page_offset = -bytes_needed;
+       skb_fill_page_desc(nskb, 0,
+ skb_shinfo(skb)->frags[j].page.p,
+ skb_shinfo(skb)->frags[j].page_offset + remaining_page_offset,
+                          skb_shinfo(skb)->frags[j].size - remaining_page_offset);
+
+       skb_frag_ref(skb, j);
+
+       /* Fill the rest of the frags */
+       for (i = 1; i < k; i++) {
+               j = skb_shinfo(skb)->nr_frags - k + i;
+
+               skb_fill_page_desc(nskb, i,
+ skb_shinfo(skb)->frags[j].page.p,
+ skb_shinfo(skb)->frags[j].page_offset,
+ skb_shinfo(skb)->frags[j].size);
+               skb_frag_ref(skb, j);
+       }
+       skb_shinfo(nskb)->nr_frags = k;
+
+       remaining_headlen = remaining - skb->data_len;
+
+       /* headlen contains remaining data? */
+       if (remaining_headlen > 0)
+               skb_copy_bits(skb, skb->len - remaining, nskb->data + headlen,
+                             remaining_headlen);
+       nskb->len = remaining + headlen;
+       nskb->data_len =  payload_len - sizeof(struct udphdr) +
+               max_t(int, 0, remaining_headlen);
+       nskb->protocol = skb->protocol;
+       if (nskb->protocol == htons(ETH_P_IP)) {
+               ip_hdr(nskb)->id = htons(ntohs(ip_hdr(nskb)->id) +
+ skb_shinfo(skb)->gso_segs);
+               ip_hdr(nskb)->tot_len =
+                       htons(payload_len + sizeof(struct iphdr));
+       } else {
+               ipv6_hdr(nskb)->payload_len = htons(payload_len);
+       }
+       udp_hdr(nskb)->len = htons(payload_len);
+       skb_shinfo(nskb)->gso_size = 0;
+       nskb->ip_summed = skb->ip_summed;
+       nskb->csum_start = skb->csum_start;
+       nskb->csum_offset = skb->csum_offset;
+       nskb->queue_mapping = skb->queue_mapping;
+}
+
+/* might send skbs and update wqe and pi */
+struct sk_buff *mlx5e_udp_gso_handle_tx_skb(struct net_device *netdev,
+                                           struct mlx5e_txqsq *sq,
+                                           struct sk_buff *skb,
+                                           struct mlx5e_tx_wqe **wqe,
+                                           u16 *pi)
+{
+       int payload_len = skb_shinfo(skb)->gso_size + sizeof(struct udphdr); +       int headlen = skb_transport_offset(skb) + sizeof(struct udphdr); +       int remaining = (skb->len - headlen) % skb_shinfo(skb)->gso_size;
+       struct sk_buff *nskb;
+
+       if (skb->protocol == htons(ETH_P_IP))
+               ip_hdr(skb)->tot_len = htons(payload_len + sizeof(struct iphdr));
+       else
+               ipv6_hdr(skb)->payload_len = htons(payload_len);
+       udp_hdr(skb)->len = htons(payload_len);
+       if (!remaining)
+               return skb;
+
+       nskb = alloc_skb(max_t(int, headlen, headlen + remaining - skb->data_len), GFP_ATOMIC);
+       if (unlikely(!nskb)) {
+               sq->stats->dropped++;
+               return NULL;
+       }
+
+       mlx5e_udp_gso_prepare_last_skb(skb, nskb, remaining);
+
+       skb_shinfo(skb)->gso_segs--;
+       pskb_trim(skb, skb->len - remaining);
+       mlx5e_sq_xmit(sq, skb, *wqe, *pi);
+       mlx5e_sq_fetch_wqe(sq, wqe, pi);
+       return nskb;
+}
The device driver seems to be implementing the packet split here
similar to NETIF_F_GSO_PARTIAL. When advertising the right flag, the
stack should be able to do that for you and pass two packets to the
driver.

We've totally missed that!
Thank you. I'll fix and resubmit.

I've noticed that we could get cleaner code in our driver if we remove these two lines from net/ipv4/udp_offload.c:
if (skb_is_gso(segs))
             mss *= skb_shinfo(segs)->gso_segs;

I think that this is correct in case of GSO_PARTIAL segmentation for the following reasons: 1. After this change the UDP payload field is consistent with the IP header payload length field. Currently, IPv4 length is 1500 and UDP total length is the full unsegmented length. 2. AFAIU, in GSO_PARTIAL no tunnel headers should be modified except the IP ID field, including the UDP length field.
What do you think?

Reply via email to