xref: /dpdk/examples/distributor/main.c (revision 59f3a8acbcdbafeebe816a26d76dfb06e6450f31)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2017 Intel Corporation
3  */
4 
5 #include <stdint.h>
6 #include <inttypes.h>
7 #include <unistd.h>
8 #include <signal.h>
9 #include <getopt.h>
10 
11 #include <rte_eal.h>
12 #include <rte_ethdev.h>
13 #include <rte_cycles.h>
14 #include <rte_malloc.h>
15 #include <rte_debug.h>
16 #include <rte_prefetch.h>
17 #include <rte_distributor.h>
18 #include <rte_pause.h>
19 #include <rte_power.h>
20 
21 #define RX_RING_SIZE 1024
22 #define TX_RING_SIZE 1024
23 #define NUM_MBUFS ((64*1024)-1)
24 #define MBUF_CACHE_SIZE 128
25 #define BURST_SIZE 64
26 #define SCHED_RX_RING_SZ 8192
27 #define SCHED_TX_RING_SZ 65536
28 #define BURST_SIZE_TX 32
29 
30 #define RTE_LOGTYPE_DISTRAPP RTE_LOGTYPE_USER1
31 
32 #define ANSI_COLOR_RED     "\x1b[31m"
33 #define ANSI_COLOR_RESET   "\x1b[0m"
34 
35 /* mask of enabled ports */
36 static uint32_t enabled_port_mask;
37 volatile uint8_t quit_signal;
38 volatile uint8_t quit_signal_rx;
39 volatile uint8_t quit_signal_dist;
40 volatile uint8_t quit_signal_work;
41 unsigned int power_lib_initialised;
42 
43 static volatile struct app_stats {
44 	struct {
45 		uint64_t rx_pkts;
46 		uint64_t returned_pkts;
47 		uint64_t enqueued_pkts;
48 		uint64_t enqdrop_pkts;
49 	} rx __rte_cache_aligned;
50 	int pad1 __rte_cache_aligned;
51 
52 	struct {
53 		uint64_t in_pkts;
54 		uint64_t ret_pkts;
55 		uint64_t sent_pkts;
56 		uint64_t enqdrop_pkts;
57 	} dist __rte_cache_aligned;
58 	int pad2 __rte_cache_aligned;
59 
60 	struct {
61 		uint64_t dequeue_pkts;
62 		uint64_t tx_pkts;
63 		uint64_t enqdrop_pkts;
64 	} tx __rte_cache_aligned;
65 	int pad3 __rte_cache_aligned;
66 
67 	uint64_t worker_pkts[64] __rte_cache_aligned;
68 
69 	int pad4 __rte_cache_aligned;
70 
71 	uint64_t worker_bursts[64][8] __rte_cache_aligned;
72 
73 	int pad5 __rte_cache_aligned;
74 
75 	uint64_t port_rx_pkts[64] __rte_cache_aligned;
76 	uint64_t port_tx_pkts[64] __rte_cache_aligned;
77 } app_stats;
78 
79 struct app_stats prev_app_stats;
80 
81 static const struct rte_eth_conf port_conf_default = {
82 	.rxmode = {
83 		.mq_mode = ETH_MQ_RX_RSS,
84 	},
85 	.txmode = {
86 		.mq_mode = ETH_MQ_TX_NONE,
87 	},
88 	.rx_adv_conf = {
89 		.rss_conf = {
90 			.rss_hf = ETH_RSS_IP | ETH_RSS_UDP |
91 				ETH_RSS_TCP | ETH_RSS_SCTP,
92 		}
93 	},
94 };
95 
96 struct output_buffer {
97 	unsigned count;
98 	struct rte_mbuf *mbufs[BURST_SIZE];
99 };
100 
101 static void print_stats(void);
102 
103 /*
104  * Initialises a given port using global settings and with the rx buffers
105  * coming from the mbuf_pool passed as parameter
106  */
107 static inline int
108 port_init(uint16_t port, struct rte_mempool *mbuf_pool)
109 {
110 	struct rte_eth_conf port_conf = port_conf_default;
111 	const uint16_t rxRings = 1, txRings = rte_lcore_count() - 1;
112 	int retval;
113 	uint16_t q;
114 	uint16_t nb_rxd = RX_RING_SIZE;
115 	uint16_t nb_txd = TX_RING_SIZE;
116 	struct rte_eth_dev_info dev_info;
117 	struct rte_eth_txconf txconf;
118 
119 	if (!rte_eth_dev_is_valid_port(port))
120 		return -1;
121 
122 	retval = rte_eth_dev_info_get(port, &dev_info);
123 	if (retval != 0) {
124 		printf("Error during getting device (port %u) info: %s\n",
125 				port, strerror(-retval));
126 		return retval;
127 	}
128 
129 	if (dev_info.tx_offload_capa & DEV_TX_OFFLOAD_MBUF_FAST_FREE)
130 		port_conf.txmode.offloads |=
131 			DEV_TX_OFFLOAD_MBUF_FAST_FREE;
132 
133 	port_conf.rx_adv_conf.rss_conf.rss_hf &=
134 		dev_info.flow_type_rss_offloads;
135 	if (port_conf.rx_adv_conf.rss_conf.rss_hf !=
136 			port_conf_default.rx_adv_conf.rss_conf.rss_hf) {
137 		printf("Port %u modified RSS hash function based on hardware support,"
138 			"requested:%#"PRIx64" configured:%#"PRIx64"\n",
139 			port,
140 			port_conf_default.rx_adv_conf.rss_conf.rss_hf,
141 			port_conf.rx_adv_conf.rss_conf.rss_hf);
142 	}
143 
144 	retval = rte_eth_dev_configure(port, rxRings, txRings, &port_conf);
145 	if (retval != 0)
146 		return retval;
147 
148 	retval = rte_eth_dev_adjust_nb_rx_tx_desc(port, &nb_rxd, &nb_txd);
149 	if (retval != 0)
150 		return retval;
151 
152 	for (q = 0; q < rxRings; q++) {
153 		retval = rte_eth_rx_queue_setup(port, q, nb_rxd,
154 						rte_eth_dev_socket_id(port),
155 						NULL, mbuf_pool);
156 		if (retval < 0)
157 			return retval;
158 	}
159 
160 	txconf = dev_info.default_txconf;
161 	txconf.offloads = port_conf.txmode.offloads;
162 	for (q = 0; q < txRings; q++) {
163 		retval = rte_eth_tx_queue_setup(port, q, nb_txd,
164 						rte_eth_dev_socket_id(port),
165 						&txconf);
166 		if (retval < 0)
167 			return retval;
168 	}
169 
170 	retval = rte_eth_dev_start(port);
171 	if (retval < 0)
172 		return retval;
173 
174 	struct rte_eth_link link;
175 	do {
176 		retval = rte_eth_link_get_nowait(port, &link);
177 		if (retval < 0) {
178 			printf("Failed link get (port %u): %s\n",
179 				port, rte_strerror(-retval));
180 			return retval;
181 		} else if (link.link_status)
182 			break;
183 
184 		printf("Waiting for Link up on port %"PRIu16"\n", port);
185 		sleep(1);
186 	} while (!link.link_status);
187 
188 	if (!link.link_status) {
189 		printf("Link down on port %"PRIu16"\n", port);
190 		return 0;
191 	}
192 
193 	struct rte_ether_addr addr;
194 	retval = rte_eth_macaddr_get(port, &addr);
195 	if (retval < 0) {
196 		printf("Failed to get MAC address (port %u): %s\n",
197 				port, rte_strerror(-retval));
198 		return retval;
199 	}
200 
201 	printf("Port %u MAC: %02"PRIx8" %02"PRIx8" %02"PRIx8
202 			" %02"PRIx8" %02"PRIx8" %02"PRIx8"\n",
203 			port, RTE_ETHER_ADDR_BYTES(&addr));
204 
205 	retval = rte_eth_promiscuous_enable(port);
206 	if (retval != 0)
207 		return retval;
208 
209 	return 0;
210 }
211 
212 struct lcore_params {
213 	unsigned worker_id;
214 	struct rte_distributor *d;
215 	struct rte_ring *rx_dist_ring;
216 	struct rte_ring *dist_tx_ring;
217 	struct rte_mempool *mem_pool;
218 };
219 
220 static int
221 lcore_rx(struct lcore_params *p)
222 {
223 	const uint16_t nb_ports = rte_eth_dev_count_avail();
224 	const int socket_id = rte_socket_id();
225 	uint16_t port;
226 	struct rte_mbuf *bufs[BURST_SIZE*2];
227 
228 	RTE_ETH_FOREACH_DEV(port) {
229 		/* skip ports that are not enabled */
230 		if ((enabled_port_mask & (1 << port)) == 0)
231 			continue;
232 
233 		if (rte_eth_dev_socket_id(port) > 0 &&
234 				rte_eth_dev_socket_id(port) != socket_id)
235 			printf("WARNING, port %u is on remote NUMA node to "
236 					"RX thread.\n\tPerformance will not "
237 					"be optimal.\n", port);
238 	}
239 
240 	printf("\nCore %u doing packet RX.\n", rte_lcore_id());
241 	port = 0;
242 	while (!quit_signal_rx) {
243 
244 		/* skip ports that are not enabled */
245 		if ((enabled_port_mask & (1 << port)) == 0) {
246 			if (++port == nb_ports)
247 				port = 0;
248 			continue;
249 		}
250 		const uint16_t nb_rx = rte_eth_rx_burst(port, 0, bufs,
251 				BURST_SIZE);
252 		if (unlikely(nb_rx == 0)) {
253 			if (++port == nb_ports)
254 				port = 0;
255 			continue;
256 		}
257 		app_stats.rx.rx_pkts += nb_rx;
258 
259 /*
260  * You can run the distributor on the rx core with this code. Returned
261  * packets are then send straight to the tx core.
262  */
263 #if 0
264 	rte_distributor_process(d, bufs, nb_rx);
265 	const uint16_t nb_ret = rte_distributor_returned_pktsd,
266 			bufs, BURST_SIZE*2);
267 
268 		app_stats.rx.returned_pkts += nb_ret;
269 		if (unlikely(nb_ret == 0)) {
270 			if (++port == nb_ports)
271 				port = 0;
272 			continue;
273 		}
274 
275 		struct rte_ring *tx_ring = p->dist_tx_ring;
276 		uint16_t sent = rte_ring_enqueue_burst(tx_ring,
277 				(void *)bufs, nb_ret, NULL);
278 #else
279 		uint16_t nb_ret = nb_rx;
280 		/*
281 		 * Swap the following two lines if you want the rx traffic
282 		 * to go directly to tx, no distribution.
283 		 */
284 		struct rte_ring *out_ring = p->rx_dist_ring;
285 		/* struct rte_ring *out_ring = p->dist_tx_ring; */
286 
287 		uint16_t sent = rte_ring_enqueue_burst(out_ring,
288 				(void *)bufs, nb_ret, NULL);
289 #endif
290 
291 		app_stats.rx.enqueued_pkts += sent;
292 		if (unlikely(sent < nb_ret)) {
293 			app_stats.rx.enqdrop_pkts +=  nb_ret - sent;
294 			RTE_LOG_DP(DEBUG, DISTRAPP,
295 				"%s:Packet loss due to full ring\n", __func__);
296 			while (sent < nb_ret)
297 				rte_pktmbuf_free(bufs[sent++]);
298 		}
299 		if (++port == nb_ports)
300 			port = 0;
301 	}
302 	if (power_lib_initialised)
303 		rte_power_exit(rte_lcore_id());
304 	/* set worker & tx threads quit flag */
305 	printf("\nCore %u exiting rx task.\n", rte_lcore_id());
306 	quit_signal = 1;
307 	return 0;
308 }
309 
310 static inline void
311 flush_one_port(struct output_buffer *outbuf, uint8_t outp)
312 {
313 	unsigned int nb_tx = rte_eth_tx_burst(outp, 0,
314 			outbuf->mbufs, outbuf->count);
315 	app_stats.tx.tx_pkts += outbuf->count;
316 
317 	if (unlikely(nb_tx < outbuf->count)) {
318 		app_stats.tx.enqdrop_pkts +=  outbuf->count - nb_tx;
319 		do {
320 			rte_pktmbuf_free(outbuf->mbufs[nb_tx]);
321 		} while (++nb_tx < outbuf->count);
322 	}
323 	outbuf->count = 0;
324 }
325 
326 static inline void
327 flush_all_ports(struct output_buffer *tx_buffers)
328 {
329 	uint16_t outp;
330 
331 	RTE_ETH_FOREACH_DEV(outp) {
332 		/* skip ports that are not enabled */
333 		if ((enabled_port_mask & (1 << outp)) == 0)
334 			continue;
335 
336 		if (tx_buffers[outp].count == 0)
337 			continue;
338 
339 		flush_one_port(&tx_buffers[outp], outp);
340 	}
341 }
342 
343 
344 
345 static int
346 lcore_distributor(struct lcore_params *p)
347 {
348 	struct rte_ring *in_r = p->rx_dist_ring;
349 	struct rte_ring *out_r = p->dist_tx_ring;
350 	struct rte_mbuf *bufs[BURST_SIZE * 4];
351 	struct rte_distributor *d = p->d;
352 
353 	printf("\nCore %u acting as distributor core.\n", rte_lcore_id());
354 	while (!quit_signal_dist) {
355 		const uint16_t nb_rx = rte_ring_dequeue_burst(in_r,
356 				(void *)bufs, BURST_SIZE*1, NULL);
357 		if (nb_rx) {
358 			app_stats.dist.in_pkts += nb_rx;
359 
360 			/* Distribute the packets */
361 			rte_distributor_process(d, bufs, nb_rx);
362 			/* Handle Returns */
363 			const uint16_t nb_ret =
364 				rte_distributor_returned_pkts(d,
365 					bufs, BURST_SIZE*2);
366 
367 			if (unlikely(nb_ret == 0))
368 				continue;
369 			app_stats.dist.ret_pkts += nb_ret;
370 
371 			uint16_t sent = rte_ring_enqueue_burst(out_r,
372 					(void *)bufs, nb_ret, NULL);
373 			app_stats.dist.sent_pkts += sent;
374 			if (unlikely(sent < nb_ret)) {
375 				app_stats.dist.enqdrop_pkts += nb_ret - sent;
376 				RTE_LOG(DEBUG, DISTRAPP,
377 					"%s:Packet loss due to full out ring\n",
378 					__func__);
379 				while (sent < nb_ret)
380 					rte_pktmbuf_free(bufs[sent++]);
381 			}
382 		}
383 	}
384 	printf("\nCore %u exiting distributor task.\n", rte_lcore_id());
385 	quit_signal_work = 1;
386 	if (power_lib_initialised)
387 		rte_power_exit(rte_lcore_id());
388 	rte_distributor_flush(d);
389 	/* Unblock any returns so workers can exit */
390 	rte_distributor_clear_returns(d);
391 	quit_signal_rx = 1;
392 	return 0;
393 }
394 
395 
396 static int
397 lcore_tx(struct rte_ring *in_r)
398 {
399 	static struct output_buffer tx_buffers[RTE_MAX_ETHPORTS];
400 	const int socket_id = rte_socket_id();
401 	uint16_t port;
402 
403 	RTE_ETH_FOREACH_DEV(port) {
404 		/* skip ports that are not enabled */
405 		if ((enabled_port_mask & (1 << port)) == 0)
406 			continue;
407 
408 		if (rte_eth_dev_socket_id(port) > 0 &&
409 				rte_eth_dev_socket_id(port) != socket_id)
410 			printf("WARNING, port %u is on remote NUMA node to "
411 					"TX thread.\n\tPerformance will not "
412 					"be optimal.\n", port);
413 	}
414 
415 	printf("\nCore %u doing packet TX.\n", rte_lcore_id());
416 	while (!quit_signal) {
417 
418 		RTE_ETH_FOREACH_DEV(port) {
419 			/* skip ports that are not enabled */
420 			if ((enabled_port_mask & (1 << port)) == 0)
421 				continue;
422 
423 			struct rte_mbuf *bufs[BURST_SIZE_TX];
424 			const uint16_t nb_rx = rte_ring_dequeue_burst(in_r,
425 					(void *)bufs, BURST_SIZE_TX, NULL);
426 			app_stats.tx.dequeue_pkts += nb_rx;
427 
428 			/* if we get no traffic, flush anything we have */
429 			if (unlikely(nb_rx == 0)) {
430 				flush_all_ports(tx_buffers);
431 				continue;
432 			}
433 
434 			/* for traffic we receive, queue it up for transmit */
435 			uint16_t i;
436 			rte_prefetch_non_temporal((void *)bufs[0]);
437 			rte_prefetch_non_temporal((void *)bufs[1]);
438 			rte_prefetch_non_temporal((void *)bufs[2]);
439 			for (i = 0; i < nb_rx; i++) {
440 				struct output_buffer *outbuf;
441 				uint8_t outp;
442 				rte_prefetch_non_temporal((void *)bufs[i + 3]);
443 				/*
444 				 * workers should update in_port to hold the
445 				 * output port value
446 				 */
447 				outp = bufs[i]->port;
448 				/* skip ports that are not enabled */
449 				if ((enabled_port_mask & (1 << outp)) == 0)
450 					continue;
451 
452 				outbuf = &tx_buffers[outp];
453 				outbuf->mbufs[outbuf->count++] = bufs[i];
454 				if (outbuf->count == BURST_SIZE_TX)
455 					flush_one_port(outbuf, outp);
456 			}
457 		}
458 	}
459 	if (power_lib_initialised)
460 		rte_power_exit(rte_lcore_id());
461 	printf("\nCore %u exiting tx task.\n", rte_lcore_id());
462 	return 0;
463 }
464 
465 static void
466 int_handler(int sig_num)
467 {
468 	printf("Exiting on signal %d\n", sig_num);
469 	/* set quit flag for rx thread to exit */
470 	quit_signal_dist = 1;
471 }
472 
473 static void
474 print_stats(void)
475 {
476 	struct rte_eth_stats eth_stats;
477 	unsigned int i, j;
478 	const unsigned int num_workers = rte_lcore_count() - 4;
479 
480 	RTE_ETH_FOREACH_DEV(i) {
481 		rte_eth_stats_get(i, &eth_stats);
482 		app_stats.port_rx_pkts[i] = eth_stats.ipackets;
483 		app_stats.port_tx_pkts[i] = eth_stats.opackets;
484 	}
485 
486 	printf("\n\nRX Thread:\n");
487 	RTE_ETH_FOREACH_DEV(i) {
488 		printf("Port %u Pktsin : %5.2f\n", i,
489 				(app_stats.port_rx_pkts[i] -
490 				prev_app_stats.port_rx_pkts[i])/1000000.0);
491 		prev_app_stats.port_rx_pkts[i] = app_stats.port_rx_pkts[i];
492 	}
493 	printf(" - Received:    %5.2f\n",
494 			(app_stats.rx.rx_pkts -
495 			prev_app_stats.rx.rx_pkts)/1000000.0);
496 	printf(" - Returned:    %5.2f\n",
497 			(app_stats.rx.returned_pkts -
498 			prev_app_stats.rx.returned_pkts)/1000000.0);
499 	printf(" - Enqueued:    %5.2f\n",
500 			(app_stats.rx.enqueued_pkts -
501 			prev_app_stats.rx.enqueued_pkts)/1000000.0);
502 	printf(" - Dropped:     %s%5.2f%s\n", ANSI_COLOR_RED,
503 			(app_stats.rx.enqdrop_pkts -
504 			prev_app_stats.rx.enqdrop_pkts)/1000000.0,
505 			ANSI_COLOR_RESET);
506 
507 	printf("Distributor thread:\n");
508 	printf(" - In:          %5.2f\n",
509 			(app_stats.dist.in_pkts -
510 			prev_app_stats.dist.in_pkts)/1000000.0);
511 	printf(" - Returned:    %5.2f\n",
512 			(app_stats.dist.ret_pkts -
513 			prev_app_stats.dist.ret_pkts)/1000000.0);
514 	printf(" - Sent:        %5.2f\n",
515 			(app_stats.dist.sent_pkts -
516 			prev_app_stats.dist.sent_pkts)/1000000.0);
517 	printf(" - Dropped      %s%5.2f%s\n", ANSI_COLOR_RED,
518 			(app_stats.dist.enqdrop_pkts -
519 			prev_app_stats.dist.enqdrop_pkts)/1000000.0,
520 			ANSI_COLOR_RESET);
521 
522 	printf("TX thread:\n");
523 	printf(" - Dequeued:    %5.2f\n",
524 			(app_stats.tx.dequeue_pkts -
525 			prev_app_stats.tx.dequeue_pkts)/1000000.0);
526 	RTE_ETH_FOREACH_DEV(i) {
527 		printf("Port %u Pktsout: %5.2f\n",
528 				i, (app_stats.port_tx_pkts[i] -
529 				prev_app_stats.port_tx_pkts[i])/1000000.0);
530 		prev_app_stats.port_tx_pkts[i] = app_stats.port_tx_pkts[i];
531 	}
532 	printf(" - Transmitted: %5.2f\n",
533 			(app_stats.tx.tx_pkts -
534 			prev_app_stats.tx.tx_pkts)/1000000.0);
535 	printf(" - Dropped:     %s%5.2f%s\n", ANSI_COLOR_RED,
536 			(app_stats.tx.enqdrop_pkts -
537 			prev_app_stats.tx.enqdrop_pkts)/1000000.0,
538 			ANSI_COLOR_RESET);
539 
540 	prev_app_stats.rx.rx_pkts = app_stats.rx.rx_pkts;
541 	prev_app_stats.rx.returned_pkts = app_stats.rx.returned_pkts;
542 	prev_app_stats.rx.enqueued_pkts = app_stats.rx.enqueued_pkts;
543 	prev_app_stats.rx.enqdrop_pkts = app_stats.rx.enqdrop_pkts;
544 	prev_app_stats.dist.in_pkts = app_stats.dist.in_pkts;
545 	prev_app_stats.dist.ret_pkts = app_stats.dist.ret_pkts;
546 	prev_app_stats.dist.sent_pkts = app_stats.dist.sent_pkts;
547 	prev_app_stats.dist.enqdrop_pkts = app_stats.dist.enqdrop_pkts;
548 	prev_app_stats.tx.dequeue_pkts = app_stats.tx.dequeue_pkts;
549 	prev_app_stats.tx.tx_pkts = app_stats.tx.tx_pkts;
550 	prev_app_stats.tx.enqdrop_pkts = app_stats.tx.enqdrop_pkts;
551 
552 	for (i = 0; i < num_workers; i++) {
553 		printf("Worker %02u Pkts: %5.2f. Bursts(1-8): ", i,
554 				(app_stats.worker_pkts[i] -
555 				prev_app_stats.worker_pkts[i])/1000000.0);
556 		for (j = 0; j < 8; j++) {
557 			printf("%"PRIu64" ", app_stats.worker_bursts[i][j]);
558 			app_stats.worker_bursts[i][j] = 0;
559 		}
560 		printf("\n");
561 		prev_app_stats.worker_pkts[i] = app_stats.worker_pkts[i];
562 	}
563 }
564 
565 static int
566 lcore_worker(struct lcore_params *p)
567 {
568 	struct rte_distributor *d = p->d;
569 	const unsigned id = p->worker_id;
570 	unsigned int num = 0;
571 	unsigned int i;
572 
573 	/*
574 	 * for single port, xor_val will be zero so we won't modify the output
575 	 * port, otherwise we send traffic from 0 to 1, 2 to 3, and vice versa
576 	 */
577 	const unsigned xor_val = (rte_eth_dev_count_avail() > 1);
578 	struct rte_mbuf *buf[8] __rte_cache_aligned;
579 
580 	for (i = 0; i < 8; i++)
581 		buf[i] = NULL;
582 
583 	app_stats.worker_pkts[p->worker_id] = 1;
584 
585 	printf("\nCore %u acting as worker core.\n", rte_lcore_id());
586 	while (!quit_signal_work) {
587 		num = rte_distributor_get_pkt(d, id, buf, buf, num);
588 		/* Do a little bit of work for each packet */
589 		for (i = 0; i < num; i++) {
590 			uint64_t t = rte_rdtsc()+100;
591 
592 			while (rte_rdtsc() < t)
593 				rte_pause();
594 			buf[i]->port ^= xor_val;
595 		}
596 
597 		app_stats.worker_pkts[p->worker_id] += num;
598 		if (num > 0)
599 			app_stats.worker_bursts[p->worker_id][num-1]++;
600 	}
601 	if (power_lib_initialised)
602 		rte_power_exit(rte_lcore_id());
603 	rte_free(p);
604 	return 0;
605 }
606 
607 static int
608 init_power_library(void)
609 {
610 	int ret = 0, lcore_id;
611 	RTE_LCORE_FOREACH_WORKER(lcore_id) {
612 		/* init power management library */
613 		ret = rte_power_init(lcore_id);
614 		if (ret) {
615 			RTE_LOG(ERR, POWER,
616 				"Library initialization failed on core %u\n",
617 				lcore_id);
618 			/*
619 			 * Return on first failure, we'll fall back
620 			 * to non-power operation
621 			 */
622 			return ret;
623 		}
624 	}
625 	return ret;
626 }
627 
628 /* display usage */
629 static void
630 print_usage(const char *prgname)
631 {
632 	printf("%s [EAL options] -- -p PORTMASK\n"
633 			"  -p PORTMASK: hexadecimal bitmask of ports to configure\n",
634 			prgname);
635 }
636 
637 static int
638 parse_portmask(const char *portmask)
639 {
640 	char *end = NULL;
641 	unsigned long pm;
642 
643 	/* parse hexadecimal string */
644 	pm = strtoul(portmask, &end, 16);
645 	if ((portmask[0] == '\0') || (end == NULL) || (*end != '\0'))
646 		return 0;
647 
648 	return pm;
649 }
650 
651 /* Parse the argument given in the command line of the application */
652 static int
653 parse_args(int argc, char **argv)
654 {
655 	int opt;
656 	char **argvopt;
657 	int option_index;
658 	char *prgname = argv[0];
659 	static struct option lgopts[] = {
660 		{NULL, 0, 0, 0}
661 	};
662 
663 	argvopt = argv;
664 
665 	while ((opt = getopt_long(argc, argvopt, "p:",
666 			lgopts, &option_index)) != EOF) {
667 
668 		switch (opt) {
669 		/* portmask */
670 		case 'p':
671 			enabled_port_mask = parse_portmask(optarg);
672 			if (enabled_port_mask == 0) {
673 				printf("invalid portmask\n");
674 				print_usage(prgname);
675 				return -1;
676 			}
677 			break;
678 
679 		default:
680 			print_usage(prgname);
681 			return -1;
682 		}
683 	}
684 
685 	if (optind <= 1) {
686 		print_usage(prgname);
687 		return -1;
688 	}
689 
690 	argv[optind-1] = prgname;
691 
692 	optind = 1; /* reset getopt lib */
693 	return 0;
694 }
695 
696 /* Main function, does initialization and calls the per-lcore functions */
697 int
698 main(int argc, char *argv[])
699 {
700 	struct rte_mempool *mbuf_pool;
701 	struct rte_distributor *d;
702 	struct rte_ring *dist_tx_ring;
703 	struct rte_ring *rx_dist_ring;
704 	struct rte_power_core_capabilities lcore_cap;
705 	unsigned int lcore_id, worker_id = 0;
706 	int distr_core_id = -1, rx_core_id = -1, tx_core_id = -1;
707 	unsigned nb_ports;
708 	uint16_t portid;
709 	uint16_t nb_ports_available;
710 	uint64_t t, freq;
711 
712 	/* catch ctrl-c so we can print on exit */
713 	signal(SIGINT, int_handler);
714 
715 	/* init EAL */
716 	int ret = rte_eal_init(argc, argv);
717 	if (ret < 0)
718 		rte_exit(EXIT_FAILURE, "Error with EAL initialization\n");
719 	argc -= ret;
720 	argv += ret;
721 
722 	/* parse application arguments (after the EAL ones) */
723 	ret = parse_args(argc, argv);
724 	if (ret < 0)
725 		rte_exit(EXIT_FAILURE, "Invalid distributor parameters\n");
726 
727 	if (rte_lcore_count() < 5)
728 		rte_exit(EXIT_FAILURE, "Error, This application needs at "
729 				"least 5 logical cores to run:\n"
730 				"1 lcore for stats (can be core 0)\n"
731 				"1 lcore for packet RX\n"
732 				"1 lcore for distribution\n"
733 				"1 lcore for packet TX\n"
734 				"and at least 1 lcore for worker threads\n");
735 
736 	if (init_power_library() == 0)
737 		power_lib_initialised = 1;
738 
739 	nb_ports = rte_eth_dev_count_avail();
740 	if (nb_ports == 0)
741 		rte_exit(EXIT_FAILURE, "Error: no ethernet ports detected\n");
742 	if (nb_ports != 1 && (nb_ports & 1))
743 		rte_exit(EXIT_FAILURE, "Error: number of ports must be even, except "
744 				"when using a single port\n");
745 
746 	mbuf_pool = rte_pktmbuf_pool_create("MBUF_POOL",
747 		NUM_MBUFS * nb_ports, MBUF_CACHE_SIZE, 0,
748 		RTE_MBUF_DEFAULT_BUF_SIZE, rte_socket_id());
749 	if (mbuf_pool == NULL)
750 		rte_exit(EXIT_FAILURE, "Cannot create mbuf pool\n");
751 	nb_ports_available = nb_ports;
752 
753 	/* initialize all ports */
754 	RTE_ETH_FOREACH_DEV(portid) {
755 		/* skip ports that are not enabled */
756 		if ((enabled_port_mask & (1 << portid)) == 0) {
757 			printf("\nSkipping disabled port %d\n", portid);
758 			nb_ports_available--;
759 			continue;
760 		}
761 		/* init port */
762 		printf("Initializing port %u... done\n", portid);
763 
764 		if (port_init(portid, mbuf_pool) != 0)
765 			rte_exit(EXIT_FAILURE, "Cannot initialize port %u\n",
766 					portid);
767 	}
768 
769 	if (!nb_ports_available) {
770 		rte_exit(EXIT_FAILURE,
771 				"All available ports are disabled. Please set portmask.\n");
772 	}
773 
774 	d = rte_distributor_create("PKT_DIST", rte_socket_id(),
775 			rte_lcore_count() - 4,
776 			RTE_DIST_ALG_BURST);
777 	if (d == NULL)
778 		rte_exit(EXIT_FAILURE, "Cannot create distributor\n");
779 
780 	/*
781 	 * scheduler ring is read by the transmitter core, and written to
782 	 * by scheduler core
783 	 */
784 	dist_tx_ring = rte_ring_create("Output_ring", SCHED_TX_RING_SZ,
785 			rte_socket_id(), RING_F_SC_DEQ | RING_F_SP_ENQ);
786 	if (dist_tx_ring == NULL)
787 		rte_exit(EXIT_FAILURE, "Cannot create output ring\n");
788 
789 	rx_dist_ring = rte_ring_create("Input_ring", SCHED_RX_RING_SZ,
790 			rte_socket_id(), RING_F_SC_DEQ | RING_F_SP_ENQ);
791 	if (rx_dist_ring == NULL)
792 		rte_exit(EXIT_FAILURE, "Cannot create output ring\n");
793 
794 	if (power_lib_initialised) {
795 		/*
796 		 * Here we'll pre-assign lcore ids to the rx, tx and
797 		 * distributor workloads if there's higher frequency
798 		 * on those cores e.g. if Turbo Boost is enabled.
799 		 * It's also worth mentioning that it will assign cores in a
800 		 * specific order, so that if there's less than three
801 		 * available, the higher frequency cores will go to the
802 		 * distributor first, then rx, then tx.
803 		 */
804 		RTE_LCORE_FOREACH_WORKER(lcore_id) {
805 
806 			rte_power_get_capabilities(lcore_id, &lcore_cap);
807 
808 			if (lcore_cap.priority != 1)
809 				continue;
810 
811 			if (distr_core_id < 0) {
812 				distr_core_id = lcore_id;
813 				printf("Distributor on priority core %d\n",
814 					lcore_id);
815 				continue;
816 			}
817 			if (rx_core_id < 0) {
818 				rx_core_id = lcore_id;
819 				printf("Rx on priority core %d\n",
820 					lcore_id);
821 				continue;
822 			}
823 			if (tx_core_id < 0) {
824 				tx_core_id = lcore_id;
825 				printf("Tx on priority core %d\n",
826 					lcore_id);
827 				continue;
828 			}
829 		}
830 	}
831 
832 	/*
833 	 * If there's any of the key workloads left without an lcore_id
834 	 * after the high performing core assignment above, pre-assign
835 	 * them here.
836 	 */
837 	RTE_LCORE_FOREACH_WORKER(lcore_id) {
838 		if (lcore_id == (unsigned int)distr_core_id ||
839 				lcore_id == (unsigned int)rx_core_id ||
840 				lcore_id == (unsigned int)tx_core_id)
841 			continue;
842 		if (distr_core_id < 0) {
843 			distr_core_id = lcore_id;
844 			printf("Distributor on core %d\n", lcore_id);
845 			continue;
846 		}
847 		if (rx_core_id < 0) {
848 			rx_core_id = lcore_id;
849 			printf("Rx on core %d\n", lcore_id);
850 			continue;
851 		}
852 		if (tx_core_id < 0) {
853 			tx_core_id = lcore_id;
854 			printf("Tx on core %d\n", lcore_id);
855 			continue;
856 		}
857 	}
858 
859 	printf(" tx id %d, dist id %d, rx id %d\n",
860 			tx_core_id,
861 			distr_core_id,
862 			rx_core_id);
863 
864 	/*
865 	 * Kick off all the worker threads first, avoiding the pre-assigned
866 	 * lcore_ids for tx, rx and distributor workloads.
867 	 */
868 	RTE_LCORE_FOREACH_WORKER(lcore_id) {
869 		if (lcore_id == (unsigned int)distr_core_id ||
870 				lcore_id == (unsigned int)rx_core_id ||
871 				lcore_id == (unsigned int)tx_core_id)
872 			continue;
873 		printf("Starting thread %d as worker, lcore_id %d\n",
874 				worker_id, lcore_id);
875 		struct lcore_params *p =
876 			rte_malloc(NULL, sizeof(*p), 0);
877 		if (!p)
878 			rte_panic("malloc failure\n");
879 		*p = (struct lcore_params){worker_id++, d, rx_dist_ring,
880 			dist_tx_ring, mbuf_pool};
881 
882 		rte_eal_remote_launch((lcore_function_t *)lcore_worker,
883 				p, lcore_id);
884 	}
885 
886 	/* Start tx core */
887 	rte_eal_remote_launch((lcore_function_t *)lcore_tx,
888 			dist_tx_ring, tx_core_id);
889 
890 	/* Start distributor core */
891 	struct lcore_params *pd =
892 		rte_malloc(NULL, sizeof(*pd), 0);
893 	if (!pd)
894 		rte_panic("malloc failure\n");
895 	*pd = (struct lcore_params){worker_id++, d,
896 		rx_dist_ring, dist_tx_ring, mbuf_pool};
897 	rte_eal_remote_launch(
898 			(lcore_function_t *)lcore_distributor,
899 			pd, distr_core_id);
900 
901 	/* Start rx core */
902 	struct lcore_params *pr =
903 		rte_malloc(NULL, sizeof(*pr), 0);
904 	if (!pr)
905 		rte_panic("malloc failure\n");
906 	*pr = (struct lcore_params){worker_id++, d, rx_dist_ring,
907 		dist_tx_ring, mbuf_pool};
908 	rte_eal_remote_launch((lcore_function_t *)lcore_rx,
909 			pr, rx_core_id);
910 
911 	freq = rte_get_timer_hz();
912 	t = rte_rdtsc() + freq;
913 	while (!quit_signal_dist) {
914 		if (t < rte_rdtsc()) {
915 			print_stats();
916 			t = rte_rdtsc() + freq;
917 		}
918 		usleep(1000);
919 	}
920 
921 	RTE_LCORE_FOREACH_WORKER(lcore_id) {
922 		if (rte_eal_wait_lcore(lcore_id) < 0)
923 			return -1;
924 	}
925 
926 	print_stats();
927 
928 	rte_free(pd);
929 	rte_free(pr);
930 
931 	/* clean up the EAL */
932 	rte_eal_cleanup();
933 
934 	return 0;
935 }
936