1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
11 #include <sys/queue.h>
16 #include <rte_common.h>
17 #include <rte_byteorder.h>
19 #include <rte_memory.h>
20 #include <rte_memcpy.h>
22 #include <rte_launch.h>
23 #include <rte_atomic.h>
24 #include <rte_cycles.h>
25 #include <rte_prefetch.h>
26 #include <rte_lcore.h>
27 #include <rte_per_lcore.h>
28 #include <rte_branch_prediction.h>
29 #include <rte_interrupts.h>
30 #include <rte_random.h>
31 #include <rte_debug.h>
32 #include <rte_ether.h>
33 #include <rte_ethdev.h>
34 #include <rte_mempool.h>
36 #include <rte_malloc.h>
37 #include <rte_fbk_hash.h>
40 #define RTE_LOGTYPE_IPv4_MULTICAST RTE_LOGTYPE_USER1
44 #define MCAST_CLONE_PORTS 2
45 #define MCAST_CLONE_SEGS 2
47 #define PKT_MBUF_DATA_SIZE RTE_MBUF_DEFAULT_BUF_SIZE
48 #define NB_PKT_MBUF 8192
50 #define HDR_MBUF_DATA_SIZE (2 * RTE_PKTMBUF_HEADROOM)
51 #define NB_HDR_MBUF (NB_PKT_MBUF * MAX_PORTS)
53 #define NB_CLONE_MBUF (NB_PKT_MBUF * MCAST_CLONE_PORTS * MCAST_CLONE_SEGS * 2)
55 /* allow max jumbo frame 9.5 KB */
56 #define JUMBO_FRAME_MAX_SIZE 0x2600
58 #define MAX_PKT_BURST 32
59 #define BURST_TX_DRAIN_US 100 /* TX drain every ~100us */
61 /* Configure how many packets ahead to prefetch, when reading packets */
62 #define PREFETCH_OFFSET 3
65 * Construct Ethernet multicast address from IPv4 multicast address.
66 * Citing RFC 1112, section 6.4:
67 * "An IP host group address is mapped to an Ethernet multicast address
68 * by placing the low-order 23-bits of the IP address into the low-order
69 * 23 bits of the Ethernet multicast address 01-00-5E-00-00-00 (hex)."
71 #define ETHER_ADDR_FOR_IPV4_MCAST(x) \
72 (rte_cpu_to_be_64(0x01005e000000ULL | ((x) & 0x7fffff)) >> 16)
75 * Configurable number of RX/TX ring descriptors
77 #define RTE_TEST_RX_DESC_DEFAULT 1024
78 #define RTE_TEST_TX_DESC_DEFAULT 1024
79 static uint16_t nb_rxd = RTE_TEST_RX_DESC_DEFAULT;
80 static uint16_t nb_txd = RTE_TEST_TX_DESC_DEFAULT;
82 /* ethernet addresses of ports */
83 static struct rte_ether_addr ports_eth_addr[MAX_PORTS];
85 /* mask of enabled ports */
86 static uint32_t enabled_port_mask = 0;
88 static uint16_t nb_ports;
90 static int rx_queue_per_lcore = 1;
94 struct rte_mbuf *m_table[MAX_PKT_BURST];
97 #define MAX_RX_QUEUE_PER_LCORE 16
98 #define MAX_TX_QUEUE_PER_PORT 16
99 struct lcore_queue_conf {
102 uint8_t rx_queue_list[MAX_RX_QUEUE_PER_LCORE];
103 uint16_t tx_queue_id[MAX_PORTS];
104 struct mbuf_table tx_mbufs[MAX_PORTS];
105 } __rte_cache_aligned;
106 static struct lcore_queue_conf lcore_queue_conf[RTE_MAX_LCORE];
108 static struct rte_eth_conf port_conf = {
110 .max_rx_pkt_len = JUMBO_FRAME_MAX_SIZE,
112 .offloads = DEV_RX_OFFLOAD_JUMBO_FRAME,
115 .mq_mode = ETH_MQ_TX_NONE,
116 .offloads = DEV_TX_OFFLOAD_MULTI_SEGS,
120 static struct rte_mempool *packet_pool, *header_pool, *clone_pool;
124 static struct rte_fbk_hash_params mcast_hash_params = {
125 .name = "MCAST_HASH",
127 .entries_per_bucket = 4,
133 struct rte_fbk_hash_table *mcast_hash = NULL;
135 struct mcast_group_params {
140 static struct mcast_group_params mcast_group_table[] = {
141 {RTE_IPV4(224,0,0,101), 0x1},
142 {RTE_IPV4(224,0,0,102), 0x2},
143 {RTE_IPV4(224,0,0,103), 0x3},
144 {RTE_IPV4(224,0,0,104), 0x4},
145 {RTE_IPV4(224,0,0,105), 0x5},
146 {RTE_IPV4(224,0,0,106), 0x6},
147 {RTE_IPV4(224,0,0,107), 0x7},
148 {RTE_IPV4(224,0,0,108), 0x8},
149 {RTE_IPV4(224,0,0,109), 0x9},
150 {RTE_IPV4(224,0,0,110), 0xA},
151 {RTE_IPV4(224,0,0,111), 0xB},
152 {RTE_IPV4(224,0,0,112), 0xC},
153 {RTE_IPV4(224,0,0,113), 0xD},
154 {RTE_IPV4(224,0,0,114), 0xE},
155 {RTE_IPV4(224,0,0,115), 0xF},
158 #define N_MCAST_GROUPS \
159 (sizeof (mcast_group_table) / sizeof (mcast_group_table[0]))
162 /* Send burst of packets on an output interface */
164 send_burst(struct lcore_queue_conf *qconf, uint16_t port)
166 struct rte_mbuf **m_table;
170 queueid = qconf->tx_queue_id[port];
171 m_table = (struct rte_mbuf **)qconf->tx_mbufs[port].m_table;
172 n = qconf->tx_mbufs[port].len;
174 ret = rte_eth_tx_burst(port, queueid, m_table, n);
175 while (unlikely (ret < n)) {
176 rte_pktmbuf_free(m_table[ret]);
180 qconf->tx_mbufs[port].len = 0;
183 /* Get number of bits set. */
184 static inline uint32_t
189 for (n = 0; v != 0; v &= v - 1, n++)
196 * Create the output multicast packet based on the given input packet.
197 * There are two approaches for creating outgoing packet, though both
198 * are based on data zero-copy idea, they differ in few details:
199 * First one creates a clone of the input packet, e.g - walk though all
200 * segments of the input packet, and for each of them create a new packet
201 * mbuf and attach that new mbuf to the segment (refer to rte_pktmbuf_clone()
202 * for more details). Then new mbuf is allocated for the packet header
203 * and is prepended to the 'clone' mbuf.
204 * Second approach doesn't make a clone, it just increment refcnt for all
205 * input packet segments. Then it allocates new mbuf for the packet header
206 * and prepends it to the input packet.
207 * Basically first approach reuses only input packet's data, but creates
208 * it's own copy of packet's metadata. Second approach reuses both input's
209 * packet data and metadata.
210 * The advantage of first approach - is that each outgoing packet has it's
211 * own copy of metadata, so we can safely modify data pointer of the
212 * input packet. That allows us to skip creation if the output packet for
213 * the last destination port, but instead modify input packet's header inplace,
214 * e.g: for N destination ports we need to invoke mcast_out_pkt (N-1) times.
215 * The advantage of second approach - less work for each outgoing packet,
216 * e.g: we skip "clone" operation completely. Though it comes with a price -
217 * input packet's metadata has to be intact. So for N destination ports we
218 * need to invoke mcast_out_pkt N times.
219 * So for small number of outgoing ports (and segments in the input packet)
220 * first approach will be faster.
221 * As number of outgoing ports (and/or input segments) will grow,
222 * second way will become more preferable.
227 * Control which of the two approaches described above should be used:
228 * - 0 - use second approach:
229 * Don't "clone" input packet.
230 * Prepend new header directly to the input packet
231 * - 1 - use first approach:
232 * Make a "clone" of input packet first.
233 * Prepend new header to the clone of the input packet
235 * - The pointer to the new outgoing packet.
236 * - NULL if operation failed.
238 static inline struct rte_mbuf *
239 mcast_out_pkt(struct rte_mbuf *pkt, int use_clone)
241 struct rte_mbuf *hdr;
243 /* Create new mbuf for the header. */
244 if (unlikely ((hdr = rte_pktmbuf_alloc(header_pool)) == NULL))
247 /* If requested, then make a new clone packet. */
248 if (use_clone != 0 &&
249 unlikely ((pkt = rte_pktmbuf_clone(pkt, clone_pool)) == NULL)) {
250 rte_pktmbuf_free(hdr);
254 /* prepend new header */
257 /* update header's fields */
258 hdr->pkt_len = (uint16_t)(hdr->data_len + pkt->pkt_len);
259 hdr->nb_segs = pkt->nb_segs + 1;
261 __rte_mbuf_sanity_check(hdr, 1);
266 * Write new Ethernet header to the outgoing packet,
267 * and put it into the outgoing queue for the given port.
270 mcast_send_pkt(struct rte_mbuf *pkt, struct rte_ether_addr *dest_addr,
271 struct lcore_queue_conf *qconf, uint16_t port)
273 struct rte_ether_hdr *ethdr;
276 /* Construct Ethernet header. */
277 ethdr = (struct rte_ether_hdr *)
278 rte_pktmbuf_prepend(pkt, (uint16_t)sizeof(*ethdr));
279 RTE_ASSERT(ethdr != NULL);
281 rte_ether_addr_copy(dest_addr, ðdr->d_addr);
282 rte_ether_addr_copy(&ports_eth_addr[port], ðdr->s_addr);
283 ethdr->ether_type = rte_be_to_cpu_16(RTE_ETHER_TYPE_IPV4);
285 /* Put new packet into the output queue */
286 len = qconf->tx_mbufs[port].len;
287 qconf->tx_mbufs[port].m_table[len] = pkt;
288 qconf->tx_mbufs[port].len = ++len;
290 /* Transmit packets */
291 if (unlikely(MAX_PKT_BURST == len))
292 send_burst(qconf, port);
295 /* Multicast forward of the input packet */
297 mcast_forward(struct rte_mbuf *m, struct lcore_queue_conf *qconf)
300 struct rte_ipv4_hdr *iphdr;
301 uint32_t dest_addr, port_mask, port_num, use_clone;
306 struct rte_ether_addr as_addr;
309 /* Remove the Ethernet header from the input packet */
310 iphdr = (struct rte_ipv4_hdr *)
311 rte_pktmbuf_adj(m, (uint16_t)sizeof(struct rte_ether_hdr));
312 RTE_ASSERT(iphdr != NULL);
314 dest_addr = rte_be_to_cpu_32(iphdr->dst_addr);
317 * Check that it is a valid multicast address and
318 * we have some active ports assigned to it.
320 if (!RTE_IS_IPV4_MCAST(dest_addr) ||
321 (hash = rte_fbk_hash_lookup(mcast_hash, dest_addr)) <= 0 ||
322 (port_mask = hash & enabled_port_mask) == 0) {
327 /* Calculate number of destination ports. */
328 port_num = bitcnt(port_mask);
330 /* Should we use rte_pktmbuf_clone() or not. */
331 use_clone = (port_num <= MCAST_CLONE_PORTS &&
332 m->nb_segs <= MCAST_CLONE_SEGS);
334 /* Mark all packet's segments as referenced port_num times */
336 rte_pktmbuf_refcnt_update(m, (uint16_t)port_num);
338 /* construct destination ethernet address */
339 dst_eth_addr.as_int = ETHER_ADDR_FOR_IPV4_MCAST(dest_addr);
341 for (port = 0; use_clone != port_mask; port_mask >>= 1, port++) {
343 /* Prepare output packet and send it out. */
344 if ((port_mask & 1) != 0) {
345 if (likely ((mc = mcast_out_pkt(m, use_clone)) != NULL))
346 mcast_send_pkt(mc, &dst_eth_addr.as_addr,
348 else if (use_clone == 0)
354 * If we making clone packets, then, for the last destination port,
355 * we can overwrite input packet's metadata.
358 mcast_send_pkt(m, &dst_eth_addr.as_addr, qconf, port);
363 /* Send burst of outgoing packet, if timeout expires. */
365 send_timeout_burst(struct lcore_queue_conf *qconf)
369 const uint64_t drain_tsc = (rte_get_tsc_hz() + US_PER_S - 1) / US_PER_S * BURST_TX_DRAIN_US;
371 cur_tsc = rte_rdtsc();
372 if (likely (cur_tsc < qconf->tx_tsc + drain_tsc))
375 for (portid = 0; portid < MAX_PORTS; portid++) {
376 if (qconf->tx_mbufs[portid].len != 0)
377 send_burst(qconf, portid);
379 qconf->tx_tsc = cur_tsc;
382 /* main processing loop */
384 main_loop(__rte_unused void *dummy)
386 struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
390 struct lcore_queue_conf *qconf;
392 lcore_id = rte_lcore_id();
393 qconf = &lcore_queue_conf[lcore_id];
396 if (qconf->n_rx_queue == 0) {
397 RTE_LOG(INFO, IPv4_MULTICAST, "lcore %u has nothing to do\n",
402 RTE_LOG(INFO, IPv4_MULTICAST, "entering main loop on lcore %u\n",
405 for (i = 0; i < qconf->n_rx_queue; i++) {
407 portid = qconf->rx_queue_list[i];
408 RTE_LOG(INFO, IPv4_MULTICAST, " -- lcoreid=%u portid=%d\n",
415 * Read packet from RX queues
417 for (i = 0; i < qconf->n_rx_queue; i++) {
419 portid = qconf->rx_queue_list[i];
420 nb_rx = rte_eth_rx_burst(portid, 0, pkts_burst,
423 /* Prefetch first packets */
424 for (j = 0; j < PREFETCH_OFFSET && j < nb_rx; j++) {
425 rte_prefetch0(rte_pktmbuf_mtod(
426 pkts_burst[j], void *));
429 /* Prefetch and forward already prefetched packets */
430 for (j = 0; j < (nb_rx - PREFETCH_OFFSET); j++) {
431 rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[
432 j + PREFETCH_OFFSET], void *));
433 mcast_forward(pkts_burst[j], qconf);
436 /* Forward remaining prefetched packets */
437 for (; j < nb_rx; j++) {
438 mcast_forward(pkts_burst[j], qconf);
442 /* Send out packets from TX queues */
443 send_timeout_burst(qconf);
449 print_usage(const char *prgname)
451 printf("%s [EAL options] -- -p PORTMASK [-q NQ]\n"
452 " -p PORTMASK: hexadecimal bitmask of ports to configure\n"
453 " -q NQ: number of queue (=ports) per lcore (default is 1)\n",
458 parse_portmask(const char *portmask)
463 /* parse hexadecimal string */
464 pm = strtoul(portmask, &end, 16);
465 if ((portmask[0] == '\0') || (end == NULL) || (*end != '\0'))
472 parse_nqueue(const char *q_arg)
477 /* parse numerical string */
479 n = strtoul(q_arg, &end, 0);
480 if (errno != 0 || end == NULL || *end != '\0' ||
481 n == 0 || n >= MAX_RX_QUEUE_PER_LCORE)
487 /* Parse the argument given in the command line of the application */
489 parse_args(int argc, char **argv)
494 char *prgname = argv[0];
495 static struct option lgopts[] = {
501 while ((opt = getopt_long(argc, argvopt, "p:q:",
502 lgopts, &option_index)) != EOF) {
507 enabled_port_mask = parse_portmask(optarg);
508 if (enabled_port_mask == 0) {
509 printf("invalid portmask\n");
510 print_usage(prgname);
517 rx_queue_per_lcore = parse_nqueue(optarg);
518 if (rx_queue_per_lcore < 0) {
519 printf("invalid queue number\n");
520 print_usage(prgname);
526 print_usage(prgname);
532 argv[optind-1] = prgname;
535 optind = 1; /* reset getopt lib */
540 print_ethaddr(const char *name, struct rte_ether_addr *eth_addr)
542 char buf[RTE_ETHER_ADDR_FMT_SIZE];
543 rte_ether_format_addr(buf, RTE_ETHER_ADDR_FMT_SIZE, eth_addr);
544 printf("%s%s", name, buf);
548 init_mcast_hash(void)
552 mcast_hash_params.socket_id = rte_socket_id();
553 mcast_hash = rte_fbk_hash_create(&mcast_hash_params);
554 if (mcast_hash == NULL){
558 for (i = 0; i < N_MCAST_GROUPS; i ++){
559 if (rte_fbk_hash_add_key(mcast_hash,
560 mcast_group_table[i].ip,
561 mcast_group_table[i].port_mask) < 0) {
569 /* Check the link status of all ports in up to 9s, and print them finally */
571 check_all_ports_link_status(uint32_t port_mask)
573 #define CHECK_INTERVAL 100 /* 100ms */
574 #define MAX_CHECK_TIME 90 /* 9s (90 * 100ms) in total */
576 uint8_t count, all_ports_up, print_flag = 0;
577 struct rte_eth_link link;
579 printf("\nChecking link status");
581 for (count = 0; count <= MAX_CHECK_TIME; count++) {
583 RTE_ETH_FOREACH_DEV(portid) {
584 if ((port_mask & (1 << portid)) == 0)
586 memset(&link, 0, sizeof(link));
587 rte_eth_link_get_nowait(portid, &link);
588 /* print link status if flag set */
589 if (print_flag == 1) {
590 if (link.link_status)
592 "Port%d Link Up. Speed %u Mbps - %s\n",
593 portid, link.link_speed,
594 (link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
595 ("full-duplex") : ("half-duplex\n"));
597 printf("Port %d Link Down\n", portid);
600 /* clear all_ports_up flag if any link down */
601 if (link.link_status == ETH_LINK_DOWN) {
606 /* after finally printing all link status, get out */
610 if (all_ports_up == 0) {
613 rte_delay_ms(CHECK_INTERVAL);
616 /* set the print_flag if all ports up or timeout */
617 if (all_ports_up == 1 || count == (MAX_CHECK_TIME - 1)) {
625 main(int argc, char **argv)
627 struct lcore_queue_conf *qconf;
628 struct rte_eth_dev_info dev_info;
629 struct rte_eth_txconf *txconf;
632 unsigned lcore_id = 0, rx_lcore_id = 0;
633 uint32_t n_tx_queue, nb_lcores;
637 ret = rte_eal_init(argc, argv);
639 rte_exit(EXIT_FAILURE, "Invalid EAL parameters\n");
643 /* parse application arguments (after the EAL ones) */
644 ret = parse_args(argc, argv);
646 rte_exit(EXIT_FAILURE, "Invalid IPV4_MULTICAST parameters\n");
648 /* create the mbuf pools */
649 packet_pool = rte_pktmbuf_pool_create("packet_pool", NB_PKT_MBUF, 32,
650 0, PKT_MBUF_DATA_SIZE, rte_socket_id());
652 if (packet_pool == NULL)
653 rte_exit(EXIT_FAILURE, "Cannot init packet mbuf pool\n");
655 header_pool = rte_pktmbuf_pool_create("header_pool", NB_HDR_MBUF, 32,
656 0, HDR_MBUF_DATA_SIZE, rte_socket_id());
658 if (header_pool == NULL)
659 rte_exit(EXIT_FAILURE, "Cannot init header mbuf pool\n");
661 clone_pool = rte_pktmbuf_pool_create("clone_pool", NB_CLONE_MBUF, 32,
662 0, 0, rte_socket_id());
664 if (clone_pool == NULL)
665 rte_exit(EXIT_FAILURE, "Cannot init clone mbuf pool\n");
667 nb_ports = rte_eth_dev_count_avail();
669 rte_exit(EXIT_FAILURE, "No physical ports!\n");
670 if (nb_ports > MAX_PORTS)
671 nb_ports = MAX_PORTS;
673 nb_lcores = rte_lcore_count();
675 /* initialize all ports */
676 RTE_ETH_FOREACH_DEV(portid) {
677 struct rte_eth_rxconf rxq_conf;
678 struct rte_eth_conf local_port_conf = port_conf;
680 /* skip ports that are not enabled */
681 if ((enabled_port_mask & (1 << portid)) == 0) {
682 printf("Skipping disabled port %d\n", portid);
686 qconf = &lcore_queue_conf[rx_lcore_id];
688 /* limit the frame size to the maximum supported by NIC */
689 rte_eth_dev_info_get(portid, &dev_info);
690 local_port_conf.rxmode.max_rx_pkt_len = RTE_MIN(
691 dev_info.max_rx_pktlen,
692 local_port_conf.rxmode.max_rx_pkt_len);
694 /* get the lcore_id for this port */
695 while (rte_lcore_is_enabled(rx_lcore_id) == 0 ||
696 qconf->n_rx_queue == (unsigned)rx_queue_per_lcore) {
699 qconf = &lcore_queue_conf[rx_lcore_id];
701 if (rx_lcore_id >= RTE_MAX_LCORE)
702 rte_exit(EXIT_FAILURE, "Not enough cores\n");
704 qconf->rx_queue_list[qconf->n_rx_queue] = portid;
708 printf("Initializing port %d on lcore %u... ", portid,
712 n_tx_queue = nb_lcores;
713 if (n_tx_queue > MAX_TX_QUEUE_PER_PORT)
714 n_tx_queue = MAX_TX_QUEUE_PER_PORT;
716 ret = rte_eth_dev_configure(portid, 1, (uint16_t)n_tx_queue,
719 rte_exit(EXIT_FAILURE, "Cannot configure device: err=%d, port=%d\n",
722 ret = rte_eth_dev_adjust_nb_rx_tx_desc(portid, &nb_rxd,
725 rte_exit(EXIT_FAILURE,
726 "Cannot adjust number of descriptors: err=%d, port=%d\n",
729 rte_eth_macaddr_get(portid, &ports_eth_addr[portid]);
730 print_ethaddr(" Address:", &ports_eth_addr[portid]);
733 /* init one RX queue */
735 printf("rxq=%hu ", queueid);
737 rxq_conf = dev_info.default_rxconf;
738 rxq_conf.offloads = local_port_conf.rxmode.offloads;
739 ret = rte_eth_rx_queue_setup(portid, queueid, nb_rxd,
740 rte_eth_dev_socket_id(portid),
744 rte_exit(EXIT_FAILURE, "rte_eth_tx_queue_setup: err=%d, port=%d\n",
747 /* init one TX queue per couple (lcore,port) */
750 RTE_LCORE_FOREACH(lcore_id) {
751 if (rte_lcore_is_enabled(lcore_id) == 0)
753 printf("txq=%u,%hu ", lcore_id, queueid);
756 txconf = &dev_info.default_txconf;
757 txconf->offloads = local_port_conf.txmode.offloads;
758 ret = rte_eth_tx_queue_setup(portid, queueid, nb_txd,
759 rte_lcore_to_socket_id(lcore_id), txconf);
761 rte_exit(EXIT_FAILURE, "rte_eth_tx_queue_setup: err=%d, "
762 "port=%d\n", ret, portid);
764 qconf = &lcore_queue_conf[lcore_id];
765 qconf->tx_queue_id[portid] = queueid;
768 rte_eth_allmulticast_enable(portid);
770 ret = rte_eth_dev_start(portid);
772 rte_exit(EXIT_FAILURE, "rte_eth_dev_start: err=%d, port=%d\n",
778 check_all_ports_link_status(enabled_port_mask);
780 /* initialize the multicast hash */
781 int retval = init_mcast_hash();
783 rte_exit(EXIT_FAILURE, "Cannot build the multicast hash\n");
785 /* launch per-lcore init on every lcore */
786 rte_eal_mp_remote_launch(main_loop, NULL, CALL_MASTER);
787 RTE_LCORE_FOREACH_SLAVE(lcore_id) {
788 if (rte_eal_wait_lcore(lcore_id) < 0)