Fix arp handling when master core is on socket 1
[samplevnf.git] / VNFs / DPPD-PROX / main.c
index 59a169d..5ab85d6 100644 (file)
@@ -224,6 +224,21 @@ static int chain_flag_state(struct task_args *targ, uint64_t flag, int is_set)
        return 0;
 }
 
+static int chain_flag_always_set(struct task_args *targ, uint64_t flag)
+{
+       return (!chain_flag_state(targ, flag, 0));
+}
+
+static int chain_flag_never_set(struct task_args *targ, uint64_t flag)
+{
+       return (!chain_flag_state(targ, flag, 1));
+}
+
+static int chain_flag_sometimes_set(struct task_args *targ, uint64_t flag)
+{
+       return (chain_flag_state(targ, flag, 1));
+}
+
 static void configure_if_tx_queues(struct task_args *targ, uint8_t socket)
 {
        uint8_t if_port;
@@ -247,39 +262,19 @@ static void configure_if_tx_queues(struct task_args *targ, uint8_t socket)
                        prox_port_cfg[if_port].n_txq = 1;
                        targ->tx_port_queue[i].queue = 0;
                }
-               /* Set the ETH_TXQ_FLAGS_NOREFCOUNT flag if none of
-                  the tasks up to the task transmitting to the port
-                  use refcnt. */
-               if (!chain_flag_state(targ, TASK_FEATURE_TXQ_FLAGS_REFCOUNT, 1)) {
-                       prox_port_cfg[if_port].tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOREFCOUNT;
-                       plog_info("\t\tEnabling No refcnt on port %d\n", if_port);
-               }
-               else {
-                       plog_info("\t\tRefcnt used on port %d\n", if_port);
-               }
-
                /* By default OFFLOAD is enabled, but if the whole
                   chain has NOOFFLOADS set all the way until the
                   first task that receives from a port, it will be
                   disabled for the destination port. */
-               if (!chain_flag_state(targ, TASK_FEATURE_TXQ_FLAGS_NOOFFLOADS, 0)) {
+#if RTE_VERSION < RTE_VERSION_NUM(18,8,0,1)
+               if (chain_flag_always_set(targ, TASK_FEATURE_TXQ_FLAGS_NOOFFLOADS)) {
                        prox_port_cfg[if_port].tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOOFFLOADS;
-                       plog_info("\t\tDisabling TX offloads on port %d\n", if_port);
-               } else {
-                       plog_info("\t\tEnabling TX offloads on port %d\n", if_port);
                }
-
-               /* By default NOMULTSEGS is disabled, as drivers/NIC might split packets on RX
-                  It should only be enabled when we know for sure that the RX does not split packets.
-                  Set the ETH_TXQ_FLAGS_NOMULTSEGS flag if all of the tasks up to the task
-                  transmitting to the port use no_multsegs. */
-               if (!chain_flag_state(targ, TASK_FEATURE_TXQ_FLAGS_NOMULTSEGS, 0)) {
-                       prox_port_cfg[if_port].tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOMULTSEGS;
-                       plog_info("\t\tEnabling No MultiSegs on port %d\n", if_port);
-               }
-               else {
-                       plog_info("\t\tMultiSegs used on port %d\n", if_port);
+#else
+               if (chain_flag_always_set(targ, TASK_FEATURE_TXQ_FLAGS_NOOFFLOADS)) {
+                       prox_port_cfg[if_port].requested_tx_offload &= ~(DEV_TX_OFFLOAD_IPV4_CKSUM | DEV_TX_OFFLOAD_UDP_CKSUM);
                }
+#endif
        }
 }
 
@@ -293,26 +288,26 @@ static void configure_if_rx_queues(struct task_args *targ, uint8_t socket)
                        return;
                }
 
-               PROX_PANIC(!prox_port_cfg[if_port].active, "Port %u not used, aborting...\n", if_port);
                port = &prox_port_cfg[if_port];
+               PROX_PANIC(!port->active, "Port %u not used, aborting...\n", if_port);
 
                if(port->rx_ring[0] != '\0') {
                        port->n_rxq = 0;
                }
 
-               // Force multi segment support if mbuf size is not big enough.
+               // If the mbuf size (of the rx task) is not big enough, we might receive multiple segments
                // This is usually the case when setting a big mtu size i.e. enabling jumbo frames.
+               // If the packets get transmitted, then multi segments will have to be enabled on the TX port
                uint16_t max_frame_size = port->mtu + ETHER_HDR_LEN + ETHER_CRC_LEN + 2 * PROX_VLAN_TAG_SIZE;
                if (max_frame_size + sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM > targ->mbuf_size) {
-                       targ->task_init->flag_features &= ~TASK_FEATURE_TXQ_FLAGS_NOMULTSEGS;
-                       plog_info("\t\tDisabling No MultSegs on port %u as %lu > %u\n", if_port, max_frame_size + sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM, targ->mbuf_size);
+                       targ->task_init->flag_features |= TASK_FEATURE_TXQ_FLAGS_MULTSEGS;
                }
-               targ->rx_port_queue[i].queue = prox_port_cfg[if_port].n_rxq;
+               targ->rx_port_queue[i].queue = port->n_rxq;
                port->pool[targ->rx_port_queue[i].queue] = targ->pool;
                port->pool_size[targ->rx_port_queue[i].queue] = targ->nb_mbuf - 1;
                port->n_rxq++;
 
-               int dsocket = prox_port_cfg[if_port].socket;
+               int dsocket = port->socket;
                if (dsocket != -1 && dsocket != socket) {
                        plog_warn("RX core on socket %d while device on socket %d\n", socket, dsocket);
                }
@@ -333,6 +328,62 @@ static void configure_if_queues(void)
        }
 }
 
+static void configure_tx_queue_flags(void)
+{
+       struct lcore_cfg *lconf = NULL;
+       struct task_args *targ;
+       uint8_t socket;
+       uint8_t if_port;
+
+        while (core_targ_next(&lconf, &targ, 0) == 0) {
+                socket = rte_lcore_to_socket_id(lconf->id);
+                for (uint8_t i = 0; i < targ->nb_txports; ++i) {
+                        if_port = targ->tx_port_queue[i].port;
+#if RTE_VERSION < RTE_VERSION_NUM(18,8,0,1)
+                        /* Set the ETH_TXQ_FLAGS_NOREFCOUNT flag if none of
+                        the tasks up to the task transmitting to the port
+                        use refcnt. */
+                        if (chain_flag_never_set(targ, TASK_FEATURE_TXQ_FLAGS_REFCOUNT)) {
+                                prox_port_cfg[if_port].tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOREFCOUNT;
+                        }
+#else
+                        /* Set the DEV_TX_OFFLOAD_MBUF_FAST_FREE flag if none of
+                        the tasks up to the task transmitting to the port
+                        use refcnt and per-queue all mbufs comes from the same mempool. */
+                        if (chain_flag_never_set(targ, TASK_FEATURE_TXQ_FLAGS_REFCOUNT)) {
+                                if (chain_flag_never_set(targ, TASK_FEATURE_TXQ_FLAGS_MULTIPLE_MEMPOOL))
+                                        prox_port_cfg[if_port].requested_tx_offload |= DEV_TX_OFFLOAD_MBUF_FAST_FREE;
+                        }
+#endif
+                }
+       }
+}
+
+static void configure_multi_segments(void)
+{
+       struct lcore_cfg *lconf = NULL;
+       struct task_args *targ;
+       uint8_t if_port;
+
+       while (core_targ_next(&lconf, &targ, 0) == 0) {
+               for (uint8_t i = 0; i < targ->nb_txports; ++i) {
+                       if_port = targ->tx_port_queue[i].port;
+                       // Multi segment is disabled for most tasks. It is only enabled for tasks requiring big packets.
+#if RTE_VERSION < RTE_VERSION_NUM(18,8,0,1)
+                       // We can only enable "no multi segment" if no such task exists in the chain of tasks.
+                       if (chain_flag_never_set(targ, TASK_FEATURE_TXQ_FLAGS_MULTSEGS)) {
+                               prox_port_cfg[if_port].tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOMULTSEGS;
+                       }
+#else
+                       // We enable "multi segment" if at least one task requires it in the chain of tasks.
+                       if (chain_flag_sometimes_set(targ, TASK_FEATURE_TXQ_FLAGS_MULTSEGS)) {
+                               prox_port_cfg[if_port].requested_tx_offload |= DEV_TX_OFFLOAD_MULTI_SEGS;
+                       }
+#endif
+               }
+       }
+}
+
 static const char *gen_ring_name(void)
 {
        static char retval[] = "XX";
@@ -450,7 +501,7 @@ static struct rte_ring *init_ring_between_tasks(struct lcore_cfg *lconf, struct
                        starg->ctrl_plane_ring = ring;
                }
 
-               plog_info("\t\tCore %u task %u to -> core %u task %u ctrl_ring %s %p %s\n",
+               plog_info("\t\t\tCore %u task %u to -> core %u task %u ctrl_ring %s %p %s\n",
                          lconf->id, starg->id, ct.core, ct.task, ct.type == CTRL_TYPE_PKT?
                          "pkt" : "msg", ring, ring->name);
                ris->n_ctrl_rings++;
@@ -509,6 +560,8 @@ static struct rte_ring *init_ring_between_tasks(struct lcore_cfg *lconf, struct
                PROX_ASSERT(dtarg->nb_rxrings < MAX_RINGS_PER_TASK);
                dtarg->rx_rings[dtarg->nb_rxrings] = ring;
                ++dtarg->nb_rxrings;
+               if (dtarg->nb_rxrings > 1)
+                       dtarg->task_init->flag_features |= TASK_FEATURE_TXQ_FLAGS_MULTIPLE_MEMPOOL;
        }
        dtarg->nb_slave_threads = starg->core_task_set[idx].n_elems;
        dtarg->lb_friend_core = lconf->id;
@@ -561,7 +614,7 @@ static void init_rings(void)
 
                        ct.core = lconf->id;
                        ct.task = starg->id;;
-                       struct rte_ring *tx_ring = init_ring_between_tasks(lcore_cfg, lcore_cfg[prox_cfg.master].targs, ct, 0, 0, &ris);
+                       struct rte_ring *tx_ring = init_ring_between_tasks(&lcore_cfg[prox_cfg.master], lcore_cfg[prox_cfg.master].targs, ct, 0, 0, &ris);
                }
        }
 }
@@ -571,13 +624,14 @@ static void shuffle_mempool(struct rte_mempool* mempool, uint32_t nb_mbuf)
        struct rte_mbuf** pkts = prox_zmalloc(nb_mbuf * sizeof(*pkts), rte_socket_id());
        uint64_t got = 0;
 
-       while (rte_mempool_get_bulk(mempool, (void**)(pkts + got), 1) == 0)
+       while ((got < nb_mbuf) && (rte_mempool_get_bulk(mempool, (void**)(pkts + got), 1) == 0))
                ++got;
 
+       nb_mbuf = got;
        while (got) {
                int idx;
                do {
-                       idx = rand() % nb_mbuf - 1;
+                       idx = rand() % nb_mbuf;
                } while (pkts[idx] == 0);
 
                rte_mempool_put_bulk(mempool, (void**)&pkts[idx], 1);
@@ -591,20 +645,16 @@ static void set_mbuf_size(struct task_args *targ)
 {
        /* mbuf size can be set
         *  - from config file (highest priority, overwriting any other config) - should only be used as workaround
-        *  - through each 'mode', overwriting the default mbuf_size
-        *  - defaulted to MBUF_SIZE i.e. 1518 Bytes
+        *  - defaulted to MBUF_SIZE.
         * Except if set explicitely, ensure that size is big enough for vmxnet3 driver
         */
-       if (targ->mbuf_size_set_explicitely)
+       if (targ->mbuf_size)
                return;
 
-       if (targ->task_init->mbuf_size != 0) {
-               /* mbuf_size not set through config file but set through mode */
-               targ->mbuf_size = targ->task_init->mbuf_size;
-       }
-
+       targ->mbuf_size = MBUF_SIZE;
        struct prox_port_cfg *port;
-       uint16_t max_frame_size = 0;
+       uint16_t max_frame_size = 0, min_buffer_size = 0;
+       int i40e = 0;
        for (int i = 0; i < targ->nb_rxports; i++) {
                uint8_t if_port = targ->rx_port_queue[i].port;
 
@@ -614,22 +664,24 @@ static void set_mbuf_size(struct task_args *targ)
                port = &prox_port_cfg[if_port];
                if (max_frame_size < port->mtu + ETHER_HDR_LEN + ETHER_CRC_LEN + 2 * PROX_VLAN_TAG_SIZE)
                        max_frame_size = port->mtu + ETHER_HDR_LEN + ETHER_CRC_LEN + 2 * PROX_VLAN_TAG_SIZE;
+               if (min_buffer_size < port->min_rx_bufsize)
+                       min_buffer_size = port->min_rx_bufsize;
 
-               if (strcmp(port->short_name, "vmxnet3") == 0) {
-                       if (targ->mbuf_size < MBUF_SIZE + RTE_PKTMBUF_HEADROOM)
-                               targ->mbuf_size = MBUF_SIZE + RTE_PKTMBUF_HEADROOM;
-                       if (targ->mbuf_size < max_frame_size)
-                               targ->mbuf_size = max_frame_size + RTE_PKTMBUF_HEADROOM;
-               }
+               // Check whether we receive from i40e. This driver have extra mbuf size requirements
+               if (strcmp(port->short_name, "i40e") == 0)
+                       i40e = 1;
        }
-       if (max_frame_size) {
+       if (i40e) {
                // i40e supports a maximum of 5 descriptors chained
                uint16_t required_mbuf_size = RTE_ALIGN(max_frame_size / 5, 128) + sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM;
                if (required_mbuf_size > targ->mbuf_size) {
                        targ->mbuf_size = required_mbuf_size;
-                       plog_info("\t\tSetting mbuf_size to %u to support frame_size %u (mtu %u)\n", targ->mbuf_size, max_frame_size, port->mtu);
+                       plog_info("\t\tSetting mbuf_size to %u to support frame_size %u\n", targ->mbuf_size, max_frame_size);
                }
        }
+       if (min_buffer_size > targ->mbuf_size) {
+               plog_warn("Mbuf size might be too small. This might result in packet segmentation and memory leak\n");
+       }
 
 }
 
@@ -918,14 +970,13 @@ static void init_lcores(void)
        plog_info("=== Initializing rings on cores ===\n");
        init_rings();
 
+       configure_multi_segments();
+       configure_tx_queue_flags();
+
        plog_info("=== Checking configuration consistency ===\n");
        check_cfg_consistent();
 
        plog_all_rings();
-
-       setup_all_task_structs_early_init();
-       plog_info("=== Initializing tasks ===\n");
-       setup_all_task_structs();
 }
 
 static int setup_prox(int argc, char **argv)
@@ -953,6 +1004,10 @@ static int setup_prox(int argc, char **argv)
        plog_info("=== Initializing ports ===\n");
        init_port_all();
 
+       setup_all_task_structs_early_init();
+       plog_info("=== Initializing tasks ===\n");
+       setup_all_task_structs();
+
        if (prox_cfg.logbuf_size) {
                prox_cfg.logbuf = prox_zmalloc(prox_cfg.logbuf_size, rte_socket_id());
                PROX_PANIC(prox_cfg.logbuf == NULL, "Failed to allocate memory for logbuf with size = %d\n", prox_cfg.logbuf_size);