i40e_bm: add udp checksum calculation

With this nopaxos looks better.

sims/nic/i40e_bm/i40e_bm.h

@@ -620,6 +620,9 @@ class i40e_bm : public nicbm::Runner::Device {
 // places the tcp checksum in the packet (assuming ipv4)
 void xsum_tcp(void *tcphdr, size_t l4len);
 
+// places the udpp checksum in the packet (assuming ipv4)
+void xsum_udp(void *udpphdr, size_t l4len);
+
 // calculates the full ipv4 & tcp checksum without assuming any pseudo header
 // xsums
 void xsum_tcpip_tso(void *iphdr, uint8_t iplen, uint8_t l4len, uint16_t paylen);

sims/nic/i40e_bm/i40e_lan.cc

@@ -591,6 +591,9 @@ bool lan_queue_tx::trigger_tx_packet() {
     if (l4t == I40E_TX_DESC_CMD_L4T_EOFT_TCP) {
       uint16_t tcp_off = maclen + iplen;
       xsum_tcp(pktbuf + tcp_off, tso_len - tcp_off);
+    } else if (l4t == I40E_TX_DESC_CMD_L4T_EOFT_UDP) {
+      uint16_t udp_off = maclen + iplen;
+      xsum_udp(pktbuf + udp_off, tso_len - udp_off);
     }
 
     runner->EthSend(pktbuf, tso_len);

sims/nic/i40e_bm/xsums.cc

@@ -30,6 +30,14 @@ struct rte_tcp_hdr {
   uint16_t tcp_urp;  /**< TCP urgent pointer, if any. */
 } __attribute__((packed));
 
+/* from dpdk/lib/librte_net/rte_udp.h */
+struct rte_udp_hdr {
+  uint16_t src_port;
+  uint16_t dst_port;
+  uint16_t dgram_len;
+  uint16_t dgram_cksum;
+} __attribute__((packed));
+
 /* from dpdk/lib/librte_net/rte_ip.h */
 struct ipv4_hdr {
   uint8_t version_ihl;      /**< version and header length */
@@ -106,6 +114,14 @@ static inline uint16_t rte_ipv4_phdr_cksum(const struct ipv4_hdr *ipv4_hdr) {
   return rte_raw_cksum(&psd_hdr, sizeof(psd_hdr));
 }
 
+void xsum_udp(void *udphdr, size_t l4_len) {
+  struct rte_udp_hdr *udph = reinterpret_cast<struct rte_udp_hdr *>(udphdr);
+  uint32_t cksum = rte_raw_cksum(udphdr, l4_len);
+  cksum = ((cksum & 0xffff0000) >> 16) + (cksum & 0xffff);
+  cksum = (~cksum) & 0xffff;
+  udph->dgram_cksum = cksum;
+}
+
 void xsum_tcp(void *tcphdr, size_t l4_len) {
   struct rte_tcp_hdr *tcph = reinterpret_cast<struct rte_tcp_hdr *>(tcphdr);
   uint32_t cksum = rte_raw_cksum(tcphdr, l4_len);