xref: /dpdk/app/test-pmd/txonly.c (revision bb85a78d)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2014 Intel Corporation
3  */
4 
5 #include <stdarg.h>
6 #include <string.h>
7 #include <stdio.h>
8 #include <errno.h>
9 #include <stdint.h>
10 #include <unistd.h>
11 #include <inttypes.h>
12 
13 #include <sys/queue.h>
14 #include <sys/stat.h>
15 
16 #include <rte_common.h>
17 #include <rte_byteorder.h>
18 #include <rte_log.h>
19 #include <rte_debug.h>
20 #include <rte_cycles.h>
21 #include <rte_memory.h>
22 #include <rte_memcpy.h>
23 #include <rte_launch.h>
24 #include <rte_eal.h>
25 #include <rte_per_lcore.h>
26 #include <rte_lcore.h>
27 #include <rte_atomic.h>
28 #include <rte_branch_prediction.h>
29 #include <rte_mempool.h>
30 #include <rte_mbuf.h>
31 #include <rte_interrupts.h>
32 #include <rte_pci.h>
33 #include <rte_ether.h>
34 #include <rte_ethdev.h>
35 #include <rte_ip.h>
36 #include <rte_tcp.h>
37 #include <rte_udp.h>
38 #include <rte_string_fns.h>
39 #include <rte_flow.h>
40 
41 #include "testpmd.h"
42 
43 struct tx_timestamp {
44 	rte_be32_t signature;
45 	rte_be16_t pkt_idx;
46 	rte_be16_t queue_idx;
47 	rte_be64_t ts;
48 };
49 
50 /* use RFC863 Discard Protocol */
51 uint16_t tx_udp_src_port = 9;
52 uint16_t tx_udp_dst_port = 9;
53 
54 /* use RFC5735 / RFC2544 reserved network test addresses */
55 uint32_t tx_ip_src_addr = (198U << 24) | (18 << 16) | (0 << 8) | 1;
56 uint32_t tx_ip_dst_addr = (198U << 24) | (18 << 16) | (0 << 8) | 2;
57 
58 #define IP_DEFTTL  64   /* from RFC 1340. */
59 
60 static struct rte_ipv4_hdr pkt_ip_hdr; /**< IP header of transmitted packets. */
61 RTE_DEFINE_PER_LCORE(uint8_t, _ip_var); /**< IP address variation */
62 static struct rte_udp_hdr pkt_udp_hdr; /**< UDP header of tx packets. */
63 RTE_DEFINE_PER_LCORE(uint64_t, timestamp_qskew);
64 					/**< Timestamp offset per queue */
65 RTE_DEFINE_PER_LCORE(uint32_t, timestamp_idone); /**< Timestamp init done. */
66 
67 static uint64_t timestamp_mask; /**< Timestamp dynamic flag mask */
68 static int32_t timestamp_off; /**< Timestamp dynamic field offset */
69 static bool timestamp_enable; /**< Timestamp enable */
70 static uint32_t timestamp_init_req; /**< Timestamp initialization request. */
71 static uint64_t timestamp_initial[RTE_MAX_ETHPORTS];
72 
73 static void
74 copy_buf_to_pkt_segs(void* buf, unsigned len, struct rte_mbuf *pkt,
75 		     unsigned offset)
76 {
77 	struct rte_mbuf *seg;
78 	void *seg_buf;
79 	unsigned copy_len;
80 
81 	seg = pkt;
82 	while (offset >= seg->data_len) {
83 		offset -= seg->data_len;
84 		seg = seg->next;
85 	}
86 	copy_len = seg->data_len - offset;
87 	seg_buf = rte_pktmbuf_mtod_offset(seg, char *, offset);
88 	while (len > copy_len) {
89 		rte_memcpy(seg_buf, buf, (size_t) copy_len);
90 		len -= copy_len;
91 		buf = ((char*) buf + copy_len);
92 		seg = seg->next;
93 		seg_buf = rte_pktmbuf_mtod(seg, char *);
94 		copy_len = seg->data_len;
95 	}
96 	rte_memcpy(seg_buf, buf, (size_t) len);
97 }
98 
99 static inline void
100 copy_buf_to_pkt(void* buf, unsigned len, struct rte_mbuf *pkt, unsigned offset)
101 {
102 	if (offset + len <= pkt->data_len) {
103 		rte_memcpy(rte_pktmbuf_mtod_offset(pkt, char *, offset),
104 			buf, (size_t) len);
105 		return;
106 	}
107 	copy_buf_to_pkt_segs(buf, len, pkt, offset);
108 }
109 
110 static void
111 setup_pkt_udp_ip_headers(struct rte_ipv4_hdr *ip_hdr,
112 			 struct rte_udp_hdr *udp_hdr,
113 			 uint16_t pkt_data_len)
114 {
115 	uint16_t *ptr16;
116 	uint32_t ip_cksum;
117 	uint16_t pkt_len;
118 
119 	/*
120 	 * Initialize UDP header.
121 	 */
122 	pkt_len = (uint16_t) (pkt_data_len + sizeof(struct rte_udp_hdr));
123 	udp_hdr->src_port = rte_cpu_to_be_16(tx_udp_src_port);
124 	udp_hdr->dst_port = rte_cpu_to_be_16(tx_udp_dst_port);
125 	udp_hdr->dgram_len      = RTE_CPU_TO_BE_16(pkt_len);
126 	udp_hdr->dgram_cksum    = 0; /* No UDP checksum. */
127 
128 	/*
129 	 * Initialize IP header.
130 	 */
131 	pkt_len = (uint16_t) (pkt_len + sizeof(struct rte_ipv4_hdr));
132 	ip_hdr->version_ihl   = RTE_IPV4_VHL_DEF;
133 	ip_hdr->type_of_service   = 0;
134 	ip_hdr->fragment_offset = 0;
135 	ip_hdr->time_to_live   = IP_DEFTTL;
136 	ip_hdr->next_proto_id = IPPROTO_UDP;
137 	ip_hdr->packet_id = 0;
138 	ip_hdr->total_length   = RTE_CPU_TO_BE_16(pkt_len);
139 	ip_hdr->src_addr = rte_cpu_to_be_32(tx_ip_src_addr);
140 	ip_hdr->dst_addr = rte_cpu_to_be_32(tx_ip_dst_addr);
141 
142 	/*
143 	 * Compute IP header checksum.
144 	 */
145 	ptr16 = (unaligned_uint16_t*) ip_hdr;
146 	ip_cksum = 0;
147 	ip_cksum += ptr16[0]; ip_cksum += ptr16[1];
148 	ip_cksum += ptr16[2]; ip_cksum += ptr16[3];
149 	ip_cksum += ptr16[4];
150 	ip_cksum += ptr16[6]; ip_cksum += ptr16[7];
151 	ip_cksum += ptr16[8]; ip_cksum += ptr16[9];
152 
153 	/*
154 	 * Reduce 32 bit checksum to 16 bits and complement it.
155 	 */
156 	ip_cksum = ((ip_cksum & 0xFFFF0000) >> 16) +
157 		(ip_cksum & 0x0000FFFF);
158 	if (ip_cksum > 65535)
159 		ip_cksum -= 65535;
160 	ip_cksum = (~ip_cksum) & 0x0000FFFF;
161 	if (ip_cksum == 0)
162 		ip_cksum = 0xFFFF;
163 	ip_hdr->hdr_checksum = (uint16_t) ip_cksum;
164 }
165 
166 static inline void
167 update_pkt_header(struct rte_mbuf *pkt, uint32_t total_pkt_len)
168 {
169 	struct rte_ipv4_hdr *ip_hdr;
170 	struct rte_udp_hdr *udp_hdr;
171 	uint16_t pkt_data_len;
172 	uint16_t pkt_len;
173 
174 	pkt_data_len = (uint16_t) (total_pkt_len - (
175 					sizeof(struct rte_ether_hdr) +
176 					sizeof(struct rte_ipv4_hdr) +
177 					sizeof(struct rte_udp_hdr)));
178 	/* updata udp pkt length */
179 	udp_hdr = rte_pktmbuf_mtod_offset(pkt, struct rte_udp_hdr *,
180 				sizeof(struct rte_ether_hdr) +
181 				sizeof(struct rte_ipv4_hdr));
182 	pkt_len = (uint16_t) (pkt_data_len + sizeof(struct rte_udp_hdr));
183 	udp_hdr->dgram_len = RTE_CPU_TO_BE_16(pkt_len);
184 
185 	/* updata ip pkt length and csum */
186 	ip_hdr = rte_pktmbuf_mtod_offset(pkt, struct rte_ipv4_hdr *,
187 				sizeof(struct rte_ether_hdr));
188 	ip_hdr->hdr_checksum = 0;
189 	pkt_len = (uint16_t) (pkt_len + sizeof(struct rte_ipv4_hdr));
190 	ip_hdr->total_length = RTE_CPU_TO_BE_16(pkt_len);
191 	ip_hdr->hdr_checksum = rte_ipv4_cksum(ip_hdr);
192 }
193 
194 static inline bool
195 pkt_burst_prepare(struct rte_mbuf *pkt, struct rte_mempool *mbp,
196 		struct rte_ether_hdr *eth_hdr, const uint16_t vlan_tci,
197 		const uint16_t vlan_tci_outer, const uint64_t ol_flags,
198 		const uint16_t idx, const struct fwd_stream *fs)
199 {
200 	struct rte_mbuf *pkt_segs[RTE_MAX_SEGS_PER_PKT];
201 	struct rte_mbuf *pkt_seg;
202 	uint32_t nb_segs, pkt_len;
203 	uint8_t i;
204 
205 	if (unlikely(tx_pkt_split == TX_PKT_SPLIT_RND))
206 		nb_segs = rte_rand() % tx_pkt_nb_segs + 1;
207 	else
208 		nb_segs = tx_pkt_nb_segs;
209 
210 	if (nb_segs > 1) {
211 		if (rte_mempool_get_bulk(mbp, (void **)pkt_segs, nb_segs - 1))
212 			return false;
213 	}
214 
215 	rte_pktmbuf_reset_headroom(pkt);
216 	pkt->data_len = tx_pkt_seg_lengths[0];
217 	pkt->ol_flags &= EXT_ATTACHED_MBUF;
218 	pkt->ol_flags |= ol_flags;
219 	pkt->vlan_tci = vlan_tci;
220 	pkt->vlan_tci_outer = vlan_tci_outer;
221 	pkt->l2_len = sizeof(struct rte_ether_hdr);
222 	pkt->l3_len = sizeof(struct rte_ipv4_hdr);
223 
224 	pkt_len = pkt->data_len;
225 	pkt_seg = pkt;
226 	for (i = 1; i < nb_segs; i++) {
227 		pkt_seg->next = pkt_segs[i - 1];
228 		pkt_seg = pkt_seg->next;
229 		pkt_seg->data_len = tx_pkt_seg_lengths[i];
230 		pkt_len += pkt_seg->data_len;
231 	}
232 	pkt_seg->next = NULL; /* Last segment of packet. */
233 	/*
234 	 * Copy headers in first packet segment(s).
235 	 */
236 	copy_buf_to_pkt(eth_hdr, sizeof(*eth_hdr), pkt, 0);
237 	copy_buf_to_pkt(&pkt_ip_hdr, sizeof(pkt_ip_hdr), pkt,
238 			sizeof(struct rte_ether_hdr));
239 	if (txonly_multi_flow) {
240 		uint8_t  ip_var = RTE_PER_LCORE(_ip_var);
241 		struct rte_ipv4_hdr *ip_hdr;
242 		uint32_t addr;
243 
244 		ip_hdr = rte_pktmbuf_mtod_offset(pkt,
245 				struct rte_ipv4_hdr *,
246 				sizeof(struct rte_ether_hdr));
247 		/*
248 		 * Generate multiple flows by varying IP src addr. This
249 		 * enables packets are well distributed by RSS in
250 		 * receiver side if any and txonly mode can be a decent
251 		 * packet generator for developer's quick performance
252 		 * regression test.
253 		 */
254 		addr = (tx_ip_dst_addr | (ip_var++ << 8)) + rte_lcore_id();
255 		ip_hdr->src_addr = rte_cpu_to_be_32(addr);
256 		RTE_PER_LCORE(_ip_var) = ip_var;
257 	}
258 	copy_buf_to_pkt(&pkt_udp_hdr, sizeof(pkt_udp_hdr), pkt,
259 			sizeof(struct rte_ether_hdr) +
260 			sizeof(struct rte_ipv4_hdr));
261 
262 	if (unlikely(tx_pkt_split == TX_PKT_SPLIT_RND) || txonly_multi_flow)
263 		update_pkt_header(pkt, pkt_len);
264 
265 	if (unlikely(timestamp_enable)) {
266 		uint64_t skew = RTE_PER_LCORE(timestamp_qskew);
267 		struct tx_timestamp timestamp_mark;
268 
269 		if (unlikely(timestamp_init_req !=
270 				RTE_PER_LCORE(timestamp_idone))) {
271 			struct rte_eth_dev_info dev_info;
272 			unsigned int txqs_n;
273 			uint64_t phase;
274 			int ret;
275 
276 			ret = eth_dev_info_get_print_err(fs->tx_port, &dev_info);
277 			if (ret != 0) {
278 				TESTPMD_LOG(ERR,
279 					"Failed to get device info for port %d,"
280 					"could not finish timestamp init",
281 					fs->tx_port);
282 				return false;
283 			}
284 			txqs_n = dev_info.nb_tx_queues;
285 			phase = tx_pkt_times_inter * fs->tx_queue /
286 					 (txqs_n ? txqs_n : 1);
287 			/*
288 			 * Initialize the scheduling time phase shift
289 			 * depending on queue index.
290 			 */
291 			skew = timestamp_initial[fs->tx_port] +
292 			       tx_pkt_times_inter + phase;
293 			RTE_PER_LCORE(timestamp_qskew) = skew;
294 			RTE_PER_LCORE(timestamp_idone) = timestamp_init_req;
295 		}
296 		timestamp_mark.pkt_idx = rte_cpu_to_be_16(idx);
297 		timestamp_mark.queue_idx = rte_cpu_to_be_16(fs->tx_queue);
298 		timestamp_mark.signature = rte_cpu_to_be_32(0xBEEFC0DE);
299 		if (unlikely(!idx)) {
300 			skew +=	tx_pkt_times_inter;
301 			pkt->ol_flags |= timestamp_mask;
302 			*RTE_MBUF_DYNFIELD
303 				(pkt, timestamp_off, uint64_t *) = skew;
304 			RTE_PER_LCORE(timestamp_qskew) = skew;
305 			timestamp_mark.ts = rte_cpu_to_be_64(skew);
306 		} else if (tx_pkt_times_intra) {
307 			skew +=	tx_pkt_times_intra;
308 			pkt->ol_flags |= timestamp_mask;
309 			*RTE_MBUF_DYNFIELD
310 				(pkt, timestamp_off, uint64_t *) = skew;
311 			RTE_PER_LCORE(timestamp_qskew) = skew;
312 			timestamp_mark.ts = rte_cpu_to_be_64(skew);
313 		} else {
314 			timestamp_mark.ts = RTE_BE64(0);
315 		}
316 		copy_buf_to_pkt(&timestamp_mark, sizeof(timestamp_mark), pkt,
317 			sizeof(struct rte_ether_hdr) +
318 			sizeof(struct rte_ipv4_hdr) +
319 			sizeof(pkt_udp_hdr));
320 	}
321 	/*
322 	 * Complete first mbuf of packet and append it to the
323 	 * burst of packets to be transmitted.
324 	 */
325 	pkt->nb_segs = nb_segs;
326 	pkt->pkt_len = pkt_len;
327 
328 	return true;
329 }
330 
331 /*
332  * Transmit a burst of multi-segments packets.
333  */
334 static void
335 pkt_burst_transmit(struct fwd_stream *fs)
336 {
337 	struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
338 	struct rte_port *txp;
339 	struct rte_mbuf *pkt;
340 	struct rte_mempool *mbp;
341 	struct rte_ether_hdr eth_hdr;
342 	uint16_t nb_tx;
343 	uint16_t nb_pkt;
344 	uint16_t vlan_tci, vlan_tci_outer;
345 	uint32_t retry;
346 	uint64_t ol_flags = 0;
347 	uint64_t tx_offloads;
348 	uint64_t start_tsc = 0;
349 
350 	get_start_cycles(&start_tsc);
351 
352 	mbp = current_fwd_lcore()->mbp;
353 	txp = &ports[fs->tx_port];
354 	tx_offloads = txp->dev_conf.txmode.offloads;
355 	vlan_tci = txp->tx_vlan_id;
356 	vlan_tci_outer = txp->tx_vlan_id_outer;
357 	if (tx_offloads	& DEV_TX_OFFLOAD_VLAN_INSERT)
358 		ol_flags = PKT_TX_VLAN_PKT;
359 	if (tx_offloads & DEV_TX_OFFLOAD_QINQ_INSERT)
360 		ol_flags |= PKT_TX_QINQ_PKT;
361 	if (tx_offloads & DEV_TX_OFFLOAD_MACSEC_INSERT)
362 		ol_flags |= PKT_TX_MACSEC;
363 
364 	/*
365 	 * Initialize Ethernet header.
366 	 */
367 	rte_ether_addr_copy(&peer_eth_addrs[fs->peer_addr], &eth_hdr.dst_addr);
368 	rte_ether_addr_copy(&ports[fs->tx_port].eth_addr, &eth_hdr.src_addr);
369 	eth_hdr.ether_type = rte_cpu_to_be_16(RTE_ETHER_TYPE_IPV4);
370 
371 	if (rte_mempool_get_bulk(mbp, (void **)pkts_burst,
372 				nb_pkt_per_burst) == 0) {
373 		for (nb_pkt = 0; nb_pkt < nb_pkt_per_burst; nb_pkt++) {
374 			if (unlikely(!pkt_burst_prepare(pkts_burst[nb_pkt], mbp,
375 							&eth_hdr, vlan_tci,
376 							vlan_tci_outer,
377 							ol_flags,
378 							nb_pkt, fs))) {
379 				rte_mempool_put_bulk(mbp,
380 						(void **)&pkts_burst[nb_pkt],
381 						nb_pkt_per_burst - nb_pkt);
382 				break;
383 			}
384 		}
385 	} else {
386 		for (nb_pkt = 0; nb_pkt < nb_pkt_per_burst; nb_pkt++) {
387 			pkt = rte_mbuf_raw_alloc(mbp);
388 			if (pkt == NULL)
389 				break;
390 			if (unlikely(!pkt_burst_prepare(pkt, mbp, &eth_hdr,
391 							vlan_tci,
392 							vlan_tci_outer,
393 							ol_flags,
394 							nb_pkt, fs))) {
395 				rte_pktmbuf_free(pkt);
396 				break;
397 			}
398 			pkts_burst[nb_pkt] = pkt;
399 		}
400 	}
401 
402 	if (nb_pkt == 0)
403 		return;
404 
405 	nb_tx = rte_eth_tx_burst(fs->tx_port, fs->tx_queue, pkts_burst, nb_pkt);
406 
407 	/*
408 	 * Retry if necessary
409 	 */
410 	if (unlikely(nb_tx < nb_pkt) && fs->retry_enabled) {
411 		retry = 0;
412 		while (nb_tx < nb_pkt && retry++ < burst_tx_retry_num) {
413 			rte_delay_us(burst_tx_delay_time);
414 			nb_tx += rte_eth_tx_burst(fs->tx_port, fs->tx_queue,
415 					&pkts_burst[nb_tx], nb_pkt - nb_tx);
416 		}
417 	}
418 	fs->tx_packets += nb_tx;
419 
420 	if (txonly_multi_flow)
421 		RTE_PER_LCORE(_ip_var) -= nb_pkt - nb_tx;
422 
423 	inc_tx_burst_stats(fs, nb_tx);
424 	if (unlikely(nb_tx < nb_pkt)) {
425 		if (verbose_level > 0 && fs->fwd_dropped == 0)
426 			printf("port %d tx_queue %d - drop "
427 			       "(nb_pkt:%u - nb_tx:%u)=%u packets\n",
428 			       fs->tx_port, fs->tx_queue,
429 			       (unsigned) nb_pkt, (unsigned) nb_tx,
430 			       (unsigned) (nb_pkt - nb_tx));
431 		fs->fwd_dropped += (nb_pkt - nb_tx);
432 		do {
433 			rte_pktmbuf_free(pkts_burst[nb_tx]);
434 		} while (++nb_tx < nb_pkt);
435 	}
436 
437 	get_end_cycles(fs, start_tsc);
438 }
439 
440 static int
441 tx_only_begin(portid_t pi)
442 {
443 	uint16_t pkt_hdr_len, pkt_data_len;
444 	int dynf;
445 
446 	pkt_hdr_len = (uint16_t)(sizeof(struct rte_ether_hdr) +
447 				 sizeof(struct rte_ipv4_hdr) +
448 				 sizeof(struct rte_udp_hdr));
449 	pkt_data_len = tx_pkt_length - pkt_hdr_len;
450 
451 	if ((tx_pkt_split == TX_PKT_SPLIT_RND || txonly_multi_flow) &&
452 	    tx_pkt_seg_lengths[0] < pkt_hdr_len) {
453 		TESTPMD_LOG(ERR,
454 			    "Random segment number or multiple flow is enabled, "
455 			    "but tx_pkt_seg_lengths[0] %u < %u (needed)\n",
456 			    tx_pkt_seg_lengths[0], pkt_hdr_len);
457 		return -EINVAL;
458 	}
459 
460 	setup_pkt_udp_ip_headers(&pkt_ip_hdr, &pkt_udp_hdr, pkt_data_len);
461 
462 	timestamp_enable = false;
463 	timestamp_mask = 0;
464 	timestamp_off = -1;
465 	RTE_PER_LCORE(timestamp_qskew) = 0;
466 	dynf = rte_mbuf_dynflag_lookup
467 				(RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME, NULL);
468 	if (dynf >= 0)
469 		timestamp_mask = 1ULL << dynf;
470 	dynf = rte_mbuf_dynfield_lookup
471 				(RTE_MBUF_DYNFIELD_TIMESTAMP_NAME, NULL);
472 	if (dynf >= 0)
473 		timestamp_off = dynf;
474 	timestamp_enable = tx_pkt_times_inter &&
475 			   timestamp_mask &&
476 			   timestamp_off >= 0 &&
477 			   !rte_eth_read_clock(pi, &timestamp_initial[pi]);
478 
479 	if (timestamp_enable) {
480 		pkt_hdr_len += sizeof(struct tx_timestamp);
481 
482 		if (tx_pkt_split == TX_PKT_SPLIT_RND) {
483 			if (tx_pkt_seg_lengths[0] < pkt_hdr_len) {
484 				TESTPMD_LOG(ERR,
485 					    "Time stamp and random segment number are enabled, "
486 					    "but tx_pkt_seg_lengths[0] %u < %u (needed)\n",
487 					    tx_pkt_seg_lengths[0], pkt_hdr_len);
488 				return -EINVAL;
489 			}
490 		} else {
491 			uint16_t total = 0;
492 			uint8_t i;
493 
494 			for (i = 0; i < tx_pkt_nb_segs; i++) {
495 				total += tx_pkt_seg_lengths[i];
496 				if (total >= pkt_hdr_len)
497 					break;
498 			}
499 
500 			if (total < pkt_hdr_len) {
501 				TESTPMD_LOG(ERR,
502 					    "Not enough Tx segment space for time stamp info, "
503 					    "total %u < %u (needed)\n",
504 					    total, pkt_hdr_len);
505 				return -EINVAL;
506 			}
507 		}
508 		timestamp_init_req++;
509 	}
510 
511 	/* Make sure all settings are visible on forwarding cores.*/
512 	rte_wmb();
513 	return 0;
514 }
515 
516 struct fwd_engine tx_only_engine = {
517 	.fwd_mode_name  = "txonly",
518 	.port_fwd_begin = tx_only_begin,
519 	.port_fwd_end   = NULL,
520 	.packet_fwd     = pkt_burst_transmit,
521 };
522