X-Git-Url: https://gerrit.opnfv.org/gerrit/gitweb?a=blobdiff_plain;ds=inline;f=VNFs%2FDPPD-PROX%2Fmain.c;h=499a1ab70bb5ce4f7fd3085113a3eec88ffd5180;hb=5d0595c955eb37af44b68098b6fd4f33f18e7755;hp=1c4dced998b67a1e5e3a7e4df9386d575a2d47c8;hpb=f608601d06f2da7dcb52f0a0da6caebd12412ad1;p=samplevnf.git diff --git a/VNFs/DPPD-PROX/main.c b/VNFs/DPPD-PROX/main.c index 1c4dced9..499a1ab7 100644 --- a/VNFs/DPPD-PROX/main.c +++ b/VNFs/DPPD-PROX/main.c @@ -94,13 +94,15 @@ static void check_mixed_normal_pipeline(void) int all_thread_nop = 1; int generic = 0; int pipeline = 0; + int l3 = 0; for (uint8_t task_id = 0; task_id < lconf->n_tasks_all; ++task_id) { struct task_args *targ = &lconf->targs[task_id]; - all_thread_nop = all_thread_nop && + l3 = !strcmp("l3", targ->sub_mode_str); + all_thread_nop = all_thread_nop && !l3 && targ->task_init->thread_x == thread_nop; pipeline = pipeline || targ->task_init->thread_x == thread_pipeline; - generic = generic || targ->task_init->thread_x == thread_generic; + generic = generic || targ->task_init->thread_x == thread_generic || l3; } PROX_PANIC(generic && pipeline, "Can't run both pipeline and normal thread on same core\n"); @@ -127,8 +129,8 @@ static void check_zero_rx(void) static void check_missing_rx(void) { - struct lcore_cfg *lconf = NULL, *rx_lconf = NULL; - struct task_args *targ, *rx_targ = NULL; + struct lcore_cfg *lconf = NULL, *rx_lconf = NULL, *tx_lconf = NULL; + struct task_args *targ, *rx_targ = NULL, *tx_targ = NULL; struct prox_port_cfg *port; uint8_t port_id, rx_port_id, ok; @@ -143,19 +145,38 @@ static void check_missing_rx(void) lconf = NULL; while (core_targ_next(&lconf, &targ, 0) == 0) { - if (strcmp(targ->task_init->sub_mode_str, "l3") != 0) + if (strcmp(targ->sub_mode_str, "l3") != 0) continue; - port = find_reachable_port(targ); - if (port == NULL) + + PROX_PANIC((targ->nb_rxports == 0) && (targ->nb_txports == 0), "L3 task must have a RX or a TX port\n"); + // If the L3 sub_mode receives from a port, check that there is at least one core/task + // transmitting to this port in L3 sub_mode + for (uint8_t i = 0; i < targ->nb_rxports; ++i) { + rx_port_id = targ->rx_port_queue[i].port; + ok = 0; + tx_lconf = NULL; + while (core_targ_next(&tx_lconf, &tx_targ, 0) == 0) { + if ((port_id = tx_targ->tx_port_queue[0].port) == OUT_DISCARD) + continue; + if ((rx_port_id == port_id) && (tx_targ->flags & TASK_ARG_L3)){ + ok = 1; + break; + } + } + PROX_PANIC(ok == 0, "RX L3 sub mode for port %d on core %d task %d, but no core/task transmitting on that port\n", rx_port_id, lconf->id, targ->id); + } + + // If the L3 sub_mode transmits to a port, check that there is at least one core/task + // receiving from that port in L3 sub_mode. + if ((port_id = targ->tx_port_queue[0].port) == OUT_DISCARD) continue; - port_id = port - prox_port_cfg; rx_lconf = NULL; ok = 0; plog_info("\tCore %d task %d transmitting to port %d in L3 mode\n", lconf->id, targ->id, port_id); while (core_targ_next(&rx_lconf, &rx_targ, 0) == 0) { for (uint8_t i = 0; i < rx_targ->nb_rxports; ++i) { rx_port_id = rx_targ->rx_port_queue[i].port; - if ((rx_port_id == port_id) && (rx_targ->task_init->flag_features & TASK_FEATURE_L3)){ + if ((rx_port_id == port_id) && (rx_targ->flags & TASK_ARG_L3)){ ok = 1; break; } @@ -203,6 +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; @@ -226,44 +262,25 @@ 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 - does not 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, 1)) { +#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 none of the tasks up to the task - transmitting to the port does not use 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 } } static void configure_if_rx_queues(struct task_args *targ, uint8_t socket) { + struct prox_port_cfg *port; for (int i = 0; i < targ->nb_rxports; i++) { uint8_t if_port = targ->rx_port_queue[i].port; @@ -271,18 +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(prox_port_cfg[if_port].rx_ring[0] != '\0') { - prox_port_cfg[if_port].n_rxq = 0; + if(port->rx_ring[0] != '\0') { + port->n_rxq = 0; } - targ->rx_port_queue[i].queue = prox_port_cfg[if_port].n_rxq; - prox_port_cfg[if_port].pool[targ->rx_port_queue[i].queue] = targ->pool; - prox_port_cfg[if_port].pool_size[targ->rx_port_queue[i].queue] = targ->nb_mbuf - 1; - prox_port_cfg[if_port].n_rxq++; + // 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_MULTSEGS; + } + 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); } @@ -298,8 +323,64 @@ static void configure_if_queues(void) while (core_targ_next(&lconf, &targ, 0) == 0) { socket = rte_lcore_to_socket_id(lconf->id); - configure_if_tx_queues(targ, socket); configure_if_rx_queues(targ, socket); + configure_if_tx_queues(targ, socket); + } +} + +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 + } } } @@ -479,6 +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; @@ -522,7 +605,7 @@ static void init_rings(void) lconf = NULL; struct prox_port_cfg *port; while (core_targ_next(&lconf, &starg, 1) == 0) { - if ((starg->task_init) && (starg->task_init->flag_features & TASK_FEATURE_L3)) { + if ((starg->task_init) && (starg->flags & TASK_ARG_L3)) { struct core_task ct; ct.core = prox_cfg.master; ct.task = 0; @@ -541,13 +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); @@ -557,6 +641,50 @@ static void shuffle_mempool(struct rte_mempool* mempool, uint32_t nb_mbuf) prox_free(pkts); } +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 + * - defaulted to MBUF_SIZE. + * Except if set explicitely, ensure that size is big enough for vmxnet3 driver + */ + if (targ->mbuf_size) + return; + + targ->mbuf_size = MBUF_SIZE; + struct prox_port_cfg *port; + 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; + + if (if_port == OUT_DISCARD) { + continue; + } + 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; + + // Check whether we receive from i40e. This driver have extra mbuf size requirements + if (strcmp(port->short_name, "i40e") == 0) + i40e = 1; + } + 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\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"); + } + +} + static void setup_mempools_unique_per_socket(void) { uint32_t flags = 0; @@ -574,11 +702,7 @@ static void setup_mempools_unique_per_socket(void) uint8_t socket = rte_lcore_to_socket_id(lconf->id); PROX_ASSERT(socket < MAX_SOCKETS); - if (targ->mbuf_size_set_explicitely) - flags = MEMPOOL_F_NO_SPREAD; - if ((!targ->mbuf_size_set_explicitely) && (targ->task_init->mbuf_size != 0)) { - targ->mbuf_size = targ->task_init->mbuf_size; - } + set_mbuf_size(targ); if (targ->rx_port_queue[0].port != OUT_DISCARD) { struct prox_port_cfg* port_cfg = &prox_port_cfg[targ->rx_port_queue[0].port]; PROX_ASSERT(targ->nb_mbuf != 0); @@ -595,10 +719,6 @@ static void setup_mempools_unique_per_socket(void) PROX_PANIC(mbuf_size[socket] != targ->mbuf_size, "all mbuf_size must have the same size if using a unique mempool per socket\n"); } - if ((!targ->mbuf_size_set_explicitely) && (strcmp(port_cfg->short_name, "vmxnet3") == 0)) { - if (mbuf_size[socket] < MBUF_SIZE + RTE_PKTMBUF_HEADROOM) - mbuf_size[socket] = MBUF_SIZE + RTE_PKTMBUF_HEADROOM; - } } } for (int i = 0 ; i < MAX_SOCKETS; i++) { @@ -647,24 +767,7 @@ static void setup_mempool_for_rx_task(struct lcore_cfg *lconf, struct task_args char memzone_name[64]; char name[64]; - /* 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 - * Except is set expliciteky, ensure that size is big enough for vmxnet3 driver - */ - if (targ->mbuf_size_set_explicitely) { - flags = MEMPOOL_F_NO_SPREAD; - /* targ->mbuf_size already set */ - } - else 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; - } - else if (strcmp(port_cfg->short_name, "vmxnet3") == 0) { - if (targ->mbuf_size < MBUF_SIZE + RTE_PKTMBUF_HEADROOM) - targ->mbuf_size = MBUF_SIZE + RTE_PKTMBUF_HEADROOM; - } + set_mbuf_size(targ); /* allocate memory pool for packets */ PROX_ASSERT(targ->nb_mbuf != 0); @@ -867,14 +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) @@ -902,6 +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);