Fix vm-bm

.env.template

@@ -8,6 +8,7 @@ NIC_PCI=0000:17:00.0
 # MAC of enp23s0f0np0 of the other node (for dpdk to assign dest MAC when inter-node) 
 # to check, unbind NIC from DPDK, then ip l
 OTHER_NODE_MAC=40:a6:b7:c3:4b:78
+TSO=0
 
 # c6525-25g nodes (Mellanox NICs don't need to be bound to vfio-pci)
 # NIC=enp65s0f0np0 

build_and_run.sh

@@ -73,5 +73,5 @@ sudo ./build/umanet \
   --eth-rx-cores $ETH_RX_CORES --eth-tx-cores $ETH_TX_CORES \
   --eth-rx-queues $ETH_RX_QUEUES --eth-tx-queues $ETH_TX_QUEUES \
   --vhost-rx-cores $VHOST_RX_CORES --vhost-tx-cores $VHOST_TX_CORES \
-  --show-dash $SHOW_DASH \
+--show-dash $SHOW_DASH --tso $TSO \
   > switch.log 2>&1

env.sh

@@ -13,6 +13,7 @@ set +a
 : "${ETH_TX_CORES:?missing ETH_TX_CORES}"
 : "${VHOST_RX_CORES:?missing VHOST_RX_CORES}"
 : "${VHOST_TX_CORES:?missing VHOST_TX_CORES}"
+: "${TSO:?missing TSO}"
 
 MAX_VM_COUNT=64
 

run.sh

@@ -36,5 +36,5 @@ sudo ./build/umanet \
   --eth-rx-cores $ETH_RX_CORES --eth-tx-cores $ETH_TX_CORES \
   --eth-rx-queues $ETH_RX_QUEUES --eth-tx-queues $ETH_TX_QUEUES \
   --vhost-rx-cores $VHOST_RX_CORES --vhost-tx-cores $VHOST_TX_CORES \
-  --show-dash $SHOW_DASH \
+  --show-dash $SHOW_DASH --tso $TSO \
   > switch.log 2>&1

src/fast/eth_rx.c

@@ -118,7 +118,8 @@ void eth_rx_loop(struct eth_rx_ctx *ctx) {
                     // NIC to VM path: clear offload flags and recalculate checksums
                     // Packets from NIC may have pseudo-checksums from sender's TX offload
                     // Virtio requires valid checksums in packet data, not offloaded
-                    fix_cksum(m);
+                    // uncomment this to enable TSO
+                    // fix_cksum(m);
 
                     vm_pkts[dst_vid][vm_cnt[dst_vid]++] = m;
                     if (vid_seen_mask & (1ULL << dst_vid))

src/fast/eth_tx.c

@@ -70,7 +70,8 @@ static inline int network_send(struct eth_tx_ctx *ctx, int tx_queue_id, unsigned
     // }
     // num = gro_num;
 
-    pkts_set_tso_flags(pkts, num);
+    // uncomment this to enable TSO
+    // pkts_set_tso_flags(pkts, num);
     int16_t ret = rte_eth_tx_burst(global->eth_port_id, tx_queue_id, pkts, num);
     if (ret < 0)
         ret = 0;

src/include/fastpath.h

@@ -25,7 +25,9 @@
 #ifndef FASTPATH_H_
 #define FASTPATH_H_
 
+#include "log.h"
 #include <rte_ether.h>
+#include <rte_icmp.h>
 #include <rte_ip.h>
 #include <rte_tcp.h>
 #include <rte_udp.h>
@@ -70,7 +72,7 @@
 
 #define GRO_MAX_FLOWS 2048
 #define GRO_MAX_ITEMS_PER_FLOW 32
-#define PKT_MTU 9000
+#define PKT_MTU 1500
 
 // tells NIC to segment TCP packets into smaller segments
 static inline void pkts_set_tso_flags(struct rte_mbuf **pkts, unsigned num) {
@@ -113,6 +115,22 @@ static inline void pkts_set_tso_flags(struct rte_mbuf **pkts, unsigned num) {
             } else if (ip->next_proto_id == IPPROTO_UDP) {
                 // UDP over IPv4 — only compute checksums, no TSO
                 m->ol_flags |= RTE_MBUF_F_TX_IPV4 | RTE_MBUF_F_TX_IP_CKSUM | RTE_MBUF_F_TX_UDP_CKSUM;
+            } else if (ip->next_proto_id == IPPROTO_ICMP) {
+                // m->ol_flags |= RTE_MBUF_F_TX_IPV4 | RTE_MBUF_F_TX_IP_CKSUM;
+                // No NIC offload exists for ICMP
+                m->ol_flags &= ~(RTE_MBUF_F_TX_IPV4 | RTE_MBUF_F_TX_IP_CKSUM | RTE_MBUF_F_TX_TCP_CKSUM |
+                                 RTE_MBUF_F_TX_UDP_CKSUM | RTE_MBUF_F_TX_TCP_SEG);
+
+                ip->hdr_checksum = 0;
+                ip->hdr_checksum = rte_ipv4_cksum(ip);
+
+                struct rte_icmp_hdr *icmp = (struct rte_icmp_hdr *)((uint8_t *)ip + m->l3_len);
+
+                uint16_t icmp_len = rte_be_to_cpu_16(ip->total_length) - m->l3_len;
+
+                icmp->icmp_cksum = 0;
+                icmp->icmp_cksum = rte_raw_cksum(icmp, icmp_len);
+                icmp->icmp_cksum = __rte_raw_cksum_reduce(icmp->icmp_cksum);
             } else {
                 // Other IPv4 protocols — just IPv4 checksum
                 m->ol_flags |= RTE_MBUF_F_TX_IPV4 | RTE_MBUF_F_TX_IP_CKSUM;

src/network/network.c

@@ -54,13 +54,12 @@ static struct rte_eth_conf port_conf = {
     .rxmode =
         {
             .mq_mode = RTE_ETH_MQ_RX_RSS,
-            .offloads = RTE_ETH_RX_OFFLOAD_IPV4_CKSUM | RTE_ETH_RX_OFFLOAD_TCP_CKSUM | RTE_ETH_RX_OFFLOAD_RSS_HASH,
+            .offloads = 0,
         },
     .txmode =
         {
             .mq_mode = RTE_ETH_MQ_TX_NONE,
-            .offloads = RTE_ETH_TX_OFFLOAD_TCP_TSO | RTE_ETH_TX_OFFLOAD_IPV4_CKSUM | RTE_ETH_TX_OFFLOAD_TCP_CKSUM |
-                        RTE_ETH_TX_OFFLOAD_MULTI_SEGS,
+            .offloads = 0,
         },
     .rx_adv_conf =
         {
@@ -112,18 +111,24 @@ int network_init() {
     rte_eth_macaddr_get(global->eth_port_id, &global->eth_addr);
     rte_eth_dev_info_get(global->eth_port_id, &eth_devinfo);
 
-    uint64_t rx_offloads = 0;
-
-    // Check if NIC supports these features
-    if (eth_devinfo.rx_offload_capa & RTE_ETH_RX_OFFLOAD_IPV4_CKSUM)
-        rx_offloads |= RTE_ETH_RX_OFFLOAD_IPV4_CKSUM;
-    if (eth_devinfo.rx_offload_capa & RTE_ETH_RX_OFFLOAD_TCP_CKSUM)
-        rx_offloads |= RTE_ETH_RX_OFFLOAD_TCP_CKSUM;
-    port_conf.rxmode.offloads = rx_offloads;
-
-    // TSO
-    eth_devinfo.default_txconf.offloads = port_conf.txmode.offloads;
-    eth_devinfo.default_rxconf.offloads = port_conf.rxmode.offloads;
+    if (config.enable_tso) {
+        port_conf.txmode.offloads = RTE_ETH_TX_OFFLOAD_TCP_TSO | RTE_ETH_TX_OFFLOAD_IPV4_CKSUM |
+                                    RTE_ETH_TX_OFFLOAD_TCP_CKSUM | RTE_ETH_TX_OFFLOAD_MULTI_SEGS;
+        port_conf.rxmode.offloads =
+            RTE_ETH_RX_OFFLOAD_IPV4_CKSUM | RTE_ETH_RX_OFFLOAD_TCP_CKSUM | RTE_ETH_RX_OFFLOAD_RSS_HASH;
+        uint64_t rx_offloads = 0;
+
+        // Check if NIC supports these features
+        if (eth_devinfo.rx_offload_capa & RTE_ETH_RX_OFFLOAD_IPV4_CKSUM)
+            rx_offloads |= RTE_ETH_RX_OFFLOAD_IPV4_CKSUM;
+        if (eth_devinfo.rx_offload_capa & RTE_ETH_RX_OFFLOAD_TCP_CKSUM)
+            rx_offloads |= RTE_ETH_RX_OFFLOAD_TCP_CKSUM;
+        port_conf.rxmode.offloads = rx_offloads;
+
+        // TSO
+        eth_devinfo.default_txconf.offloads = port_conf.txmode.offloads;
+        eth_devinfo.default_rxconf.offloads = port_conf.rxmode.offloads;
+    }
 
     if (eth_devinfo.max_rx_queues < config.eth_rx_cores || eth_devinfo.max_tx_queues < config.eth_tx_cores) {
         LOG_ERROR("Error: NIC does not support enough hw queues (rx=%u tx=%u)"
@@ -167,10 +172,12 @@ int network_init() {
         goto error_exit;
     }
 
-    // eth_devinfo.default_rxconf.offloads = 0;
+    if (!config.enable_tso) {
+        eth_devinfo.default_rxconf.offloads = 0;
 
-    /* enable per-queue checksum offload if requested */
-    // eth_devinfo.default_txconf.offloads = 0;
+        /* enable per-queue checksum offload if requested */
+        eth_devinfo.default_txconf.offloads = 0;
+    }
     // if (config.fp_xsumoffload) {
     //     uint64_t requested_offloads = RTE_ETH_TX_OFFLOAD_IPV4_CKSUM | RTE_ETH_TX_OFFLOAD_TCP_CKSUM;
     //     /* mask unsupported TX offloads (use same mask as port-level) */

src/slow/arp.c

@@ -7,7 +7,8 @@
 #include <rte_vhost.h>
 
 // only does ARP destined for dataplane, not VMs
-int process_arp_req(struct control_ctx *ctx, uint16_t vid, struct rte_mbuf *m, enum slow_src src) {
+int process_arp_req(struct control_ctx *ctx, uint16_t vid, uint16_t eth_queue_id, struct rte_mbuf *m,
+                    enum slow_src src) {
     struct rte_ether_hdr *eth = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
     struct rte_arp_hdr *arp = (struct rte_arp_hdr *)(eth + 1);
     if (arp->arp_opcode != rte_cpu_to_be_16(RTE_ARP_OP_REQUEST)) {
@@ -47,12 +48,11 @@ int process_arp_req(struct control_ctx *ctx, uint16_t vid, struct rte_mbuf *m, e
         int ret = rte_ring_enqueue_burst(global->vhost_tx_rings[vid], (void **)&m, 1, NULL);
         if (unlikely(ret == 0))
             LOG_WARN("[%d] Failed to enqueue ARP reply to vid=%d\n", ctx->core_id, vid);
+    } else if (src == SLOW_SRC_ETH) {
+        int ret = rte_ring_enqueue_burst(global->eth_tx_queue_rings[eth_queue_id], (void **)&m, 1, NULL);
+        if (unlikely(ret == 0))
+            LOG_WARN("[%d] Failed to enqueue ARP reply to eth_queue_id=%d\n", ctx->core_id, eth_queue_id);
     }
-    // else if (src == SLOW_SRC_ETH) { // handled in from_eth instead
-    //     // int ret = network_send(ctx, 1, &m);
-    //     if (unlikely(ret == 0))
-    //         LOG_WARN("[%d] Failed to send ARP reply to physical NIC\n", ctx->core_id);
-    // }
 
     rte_pktmbuf_free(m);
 

src/slow/slowpath.c

@@ -51,7 +51,7 @@ void slowpath_loop(struct control_ctx *ctx) {
                     LOG_ERROR("vdev_list->vdevs[%d] is NULL for ARP request\n", slow_msg->vid);
                     continue;
                 }
-                process_arp_req(ctx, slow_msg->vid, slow_msg->mbuf, slow_msg->src);
+                process_arp_req(ctx, slow_msg->vid, slow_msg->eth_queue_id, slow_msg->mbuf, slow_msg->src);
                 break;
 
             default:

src/slow/slowpath.h

@@ -25,7 +25,8 @@ struct slow_msg {
     struct rte_mbuf *mbuf;
 };
 
-int process_arp_req(struct control_ctx *ctx, uint16_t vid, struct rte_mbuf *m, enum slow_src src);
+int process_arp_req(struct control_ctx *ctx, uint16_t vid, uint16_t eth_queue_id, struct rte_mbuf *m,
+                    enum slow_src src);
 void slowpath_loop(struct control_ctx *ctx);
 
 void control_tty_init();

src/vhost/device.c

@@ -257,66 +257,71 @@ int register_vhost_drivers() {
             continue;
         }
 
-        // flags describe what the backend (you) and the guest agree on
-        // GUEST_TSO enabled - beneficial for traffic going through physical NIC with hardware TSO offload
-        uint64_t features = (1ULL << VIRTIO_NET_F_MTU) | (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
-                            (1ULL << VIRTIO_NET_F_CTRL_VQ) | (1ULL << VIRTIO_NET_F_CSUM) |
-                            (1ULL << VIRTIO_NET_F_GUEST_CSUM) | (1ULL << VIRTIO_NET_F_GUEST_UFO) |
-                            // (1ULL << VIRTIO_NET_F_GUEST_TSO4) | (1ULL << VIRTIO_NET_F_GUEST_TSO6) |
-                            (1ULL << VIRTIO_NET_F_HOST_TSO4) | (1ULL << VIRTIO_NET_F_HOST_TSO6);
-
-        rte_vhost_driver_enable_features(file, features);
-
-        // Allow the guest to send TSO packets to the host.
-        // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_TSO4);
-        // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_TSO6);
-
-        // Allow the host (your DPDK app) to send TSO packets to the guest.
-        // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_HOST_TSO4);
-        // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_HOST_TSO6);
-
-        // if (config.mergeable == 0) {
-        // }
-        // Allows the host to place one large packet across multiple guest RX buffers
-        // 1 packet → RX buf 0 + RX buf 1 + RX buf 2
-        // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_MRG_RXBUF);
-
-        // if (config.enable_tx_csum == 0) {
-        // }
-        // guests won’t compute checksums, host will do it
-        // guest sends packets with checksum fields = 0, host fills them later
-        // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_CSUM);
-        // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_CSUM);
-        // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_UFO);
-        // //
-        // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_ECN);
-
-        // if (config.enable_tso == 0) {
-        // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_HOST_TSO4);
-        // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_HOST_TSO6);
-        // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_TSO4);
-        // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_TSO6);
-        // // }
-
-        // - RTE_VHOST_USER_EXTBUF_SUPPORT (enables external buffer mbufs)
-        // - RTE_VHOST_USER_LINEARBUF_SUPPORT (required for external buffers)
-        // - VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD (required for zero copy tracking)
-        // uint64_t protocol_features = 0;
-        // if (rte_vhost_driver_get_protocol_features(file, &protocol_features) == 0) {
-        //     protocol_features |= (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD);
-        //     if (rte_vhost_driver_set_protocol_features(file, protocol_features) != 0) {
-        //         LOG_WARN("Failed to set INFLIGHT_SHMFD protocol feature for %s (zero copy may not work)\n", file);
-        //     } else {
-        //         LOG_INFO("Enabled INFLIGHT_SHMFD protocol feature for zero copy support (features: 0x%lx)\n",
-        //                  protocol_features);
-        //     }
-        // } else {
-        //     LOG_WARN("Failed to get protocol features for %s, trying to set INFLIGHT_SHMFD directly\n", file);
-        //     protocol_features = (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD);
-        //     if (rte_vhost_driver_set_protocol_features(file, protocol_features) != 0) {
-        //         LOG_WARN("Failed to set INFLIGHT_SHMFD protocol feature for %s (zero copy may not work)\n", file);
-        //     }
-        // }
+        if (config.enable_tso) {
+            // flags describe what the backend (you) and the guest agree on
+            // GUEST_TSO enabled - beneficial for traffic going through physical NIC with hardware TSO offload
+            uint64_t features = (1ULL << VIRTIO_NET_F_MTU) | (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
+                                (1ULL << VIRTIO_NET_F_CTRL_VQ) | (1ULL << VIRTIO_NET_F_CSUM) |
+                                (1ULL << VIRTIO_NET_F_GUEST_CSUM) | (1ULL << VIRTIO_NET_F_GUEST_UFO) |
+                                // (1ULL << VIRTIO_NET_F_GUEST_TSO4) | (1ULL << VIRTIO_NET_F_GUEST_TSO6) |
+                                (1ULL << VIRTIO_NET_F_HOST_TSO4) | (1ULL << VIRTIO_NET_F_HOST_TSO6);
+
+            rte_vhost_driver_enable_features(file, features);
+        } else {
+
+            // Allow the guest to send TSO packets to the host.
+            rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_TSO4);
+            rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_TSO6);
+
+            // Allow the host (your DPDK app) to send TSO packets to the guest.
+            rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_HOST_TSO4);
+            rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_HOST_TSO6);
+
+            // if (config.mergeable == 0) {
+            // }
+            // Allows the host to place one large packet across multiple guest RX buffers
+            // 1 packet → RX buf 0 + RX buf 1 + RX buf 2
+            rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_MRG_RXBUF);
+
+            // if (config.enable_tx_csum == 0) {
+            // }
+            // guests won’t compute checksums, host will do it
+            // guest sends packets with checksum fields = 0, host fills them later
+            rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_CSUM);
+            // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_CSUM);
+            // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_UFO);
+            // //
+            // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_ECN);
+
+            // if (config.enable_tso == 0) {
+            // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_HOST_TSO4);
+            // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_HOST_TSO6);
+            // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_TSO4);
+            // rte_vhost_driver_disable_features(file, 1ULL << VIRTIO_NET_F_GUEST_TSO6);
+            // // }
+
+            // - RTE_VHOST_USER_EXTBUF_SUPPORT (enables external buffer mbufs)
+            // - RTE_VHOST_USER_LINEARBUF_SUPPORT (required for external buffers)
+            // - VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD (required for zero copy tracking)
+            // uint64_t protocol_features = 0;
+            // if (rte_vhost_driver_get_protocol_features(file, &protocol_features) == 0) {
+            //     protocol_features |= (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD);
+            //     if (rte_vhost_driver_set_protocol_features(file, protocol_features) != 0) {
+            //         LOG_WARN("Failed to set INFLIGHT_SHMFD protocol feature for %s (zero copy may not work)\n",
+            //         file);
+            //     } else {
+            //         LOG_INFO("Enabled INFLIGHT_SHMFD protocol feature for zero copy support (features: 0x%lx)\n",
+            //                  protocol_features);
+            //     }
+            // } else {
+            //     LOG_WARN("Failed to get protocol features for %s, trying to set INFLIGHT_SHMFD directly\n", file);
+            //     protocol_features = (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD);
+            //     if (rte_vhost_driver_set_protocol_features(file, protocol_features) != 0) {
+            //         LOG_WARN("Failed to set INFLIGHT_SHMFD protocol feature for %s (zero copy may not work)\n",
+            //         file);
+            //     }
+            // }
+        }
 
         if (rte_vhost_driver_callback_register(file, &virtio_net_device_ops) != 0) {
             LOG_ERROR("Failed to register vhost driver callbacks for %s (socket %d/%d)\n", file, i, config.nb_sockets);

testing/commands.md

@@ -157,6 +157,7 @@ df -h
 systemctl status iperf
 sudo tcpdump -i br0
 sudo tcpdump -i enp23s0f0np0
+sudo tcpdump -vv -i eth1
 echo '{"vm":"vm12","throughput":12345}' | nc 192.168.100.1 9000
 
 lsof -i :9000