xref: /dpdk/app/test-pmd/txonly.c (revision 0d09cbc7)
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 /* use RFC863 Discard Protocol */
44 uint16_t tx_udp_src_port = 9;
45 uint16_t tx_udp_dst_port = 9;
46 
47 /* use RFC5735 / RFC2544 reserved network test addresses */
48 uint32_t tx_ip_src_addr = (198U << 24) | (18 << 16) | (0 << 8) | 1;
49 uint32_t tx_ip_dst_addr = (198U << 24) | (18 << 16) | (0 << 8) | 2;
50 
51 #define IP_DEFTTL  64   /* from RFC 1340. */
52 
53 static struct rte_ipv4_hdr pkt_ip_hdr; /**< IP header of transmitted packets. */
54 RTE_DEFINE_PER_LCORE(uint8_t, _ip_var); /**< IP address variation */
55 static struct rte_udp_hdr pkt_udp_hdr; /**< UDP header of tx packets. */
56 RTE_DEFINE_PER_LCORE(uint64_t, timestamp_qskew);
57 					/**< Timestamp offset per queue */
58 RTE_DEFINE_PER_LCORE(uint32_t, timestamp_idone); /**< Timestamp init done. */
59 
60 static uint64_t timestamp_mask; /**< Timestamp dynamic flag mask */
61 static int32_t timestamp_off; /**< Timestamp dynamic field offset */
62 static bool timestamp_enable; /**< Timestamp enable */
63 static uint32_t timestamp_init_req; /**< Timestamp initialization request. */
64 static uint64_t timestamp_initial[RTE_MAX_ETHPORTS];
65 
66 static void
67 copy_buf_to_pkt_segs(void* buf, unsigned len, struct rte_mbuf *pkt,
68 		     unsigned offset)
69 {
70 	struct rte_mbuf *seg;
71 	void *seg_buf;
72 	unsigned copy_len;
73 
74 	seg = pkt;
75 	while (offset >= seg->data_len) {
76 		offset -= seg->data_len;
77 		seg = seg->next;
78 	}
79 	copy_len = seg->data_len - offset;
80 	seg_buf = rte_pktmbuf_mtod_offset(seg, char *, offset);
81 	while (len > copy_len) {
82 		rte_memcpy(seg_buf, buf, (size_t) copy_len);
83 		len -= copy_len;
84 		buf = ((char*) buf + copy_len);
85 		seg = seg->next;
86 		seg_buf = rte_pktmbuf_mtod(seg, char *);
87 		copy_len = seg->data_len;
88 	}
89 	rte_memcpy(seg_buf, buf, (size_t) len);
90 }
91 
92 static inline void
93 copy_buf_to_pkt(void* buf, unsigned len, struct rte_mbuf *pkt, unsigned offset)
94 {
95 	if (offset + len <= pkt->data_len) {
96 		rte_memcpy(rte_pktmbuf_mtod_offset(pkt, char *, offset),
97 			buf, (size_t) len);
98 		return;
99 	}
100 	copy_buf_to_pkt_segs(buf, len, pkt, offset);
101 }
102 
103 static void
104 setup_pkt_udp_ip_headers(struct rte_ipv4_hdr *ip_hdr,
105 			 struct rte_udp_hdr *udp_hdr,
106 			 uint16_t pkt_data_len)
107 {
108 	uint16_t *ptr16;
109 	uint32_t ip_cksum;
110 	uint16_t pkt_len;
111 
112 	/*
113 	 * Initialize UDP header.
114 	 */
115 	pkt_len = (uint16_t) (pkt_data_len + sizeof(struct rte_udp_hdr));
116 	udp_hdr->src_port = rte_cpu_to_be_16(tx_udp_src_port);
117 	udp_hdr->dst_port = rte_cpu_to_be_16(tx_udp_dst_port);
118 	udp_hdr->dgram_len      = RTE_CPU_TO_BE_16(pkt_len);
119 	udp_hdr->dgram_cksum    = 0; /* No UDP checksum. */
120 
121 	/*
122 	 * Initialize IP header.
123 	 */
124 	pkt_len = (uint16_t) (pkt_len + sizeof(struct rte_ipv4_hdr));
125 	ip_hdr->version_ihl   = RTE_IPV4_VHL_DEF;
126 	ip_hdr->type_of_service   = 0;
127 	ip_hdr->fragment_offset = 0;
128 	ip_hdr->time_to_live   = IP_DEFTTL;
129 	ip_hdr->next_proto_id = IPPROTO_UDP;
130 	ip_hdr->packet_id = 0;
131 	ip_hdr->total_length   = RTE_CPU_TO_BE_16(pkt_len);
132 	ip_hdr->src_addr = rte_cpu_to_be_32(tx_ip_src_addr);
133 	ip_hdr->dst_addr = rte_cpu_to_be_32(tx_ip_dst_addr);
134 
135 	/*
136 	 * Compute IP header checksum.
137 	 */
138 	ptr16 = (unaligned_uint16_t*) ip_hdr;
139 	ip_cksum = 0;
140 	ip_cksum += ptr16[0]; ip_cksum += ptr16[1];
141 	ip_cksum += ptr16[2]; ip_cksum += ptr16[3];
142 	ip_cksum += ptr16[4];
143 	ip_cksum += ptr16[6]; ip_cksum += ptr16[7];
144 	ip_cksum += ptr16[8]; ip_cksum += ptr16[9];
145 
146 	/*
147 	 * Reduce 32 bit checksum to 16 bits and complement it.
148 	 */
149 	ip_cksum = ((ip_cksum & 0xFFFF0000) >> 16) +
150 		(ip_cksum & 0x0000FFFF);
151 	if (ip_cksum > 65535)
152 		ip_cksum -= 65535;
153 	ip_cksum = (~ip_cksum) & 0x0000FFFF;
154 	if (ip_cksum == 0)
155 		ip_cksum = 0xFFFF;
156 	ip_hdr->hdr_checksum = (uint16_t) ip_cksum;
157 }
158 
159 static inline bool
160 pkt_burst_prepare(struct rte_mbuf *pkt, struct rte_mempool *mbp,
161 		struct rte_ether_hdr *eth_hdr, const uint16_t vlan_tci,
162 		const uint16_t vlan_tci_outer, const uint64_t ol_flags,
163 		const uint16_t idx, const struct fwd_stream *fs)
164 {
165 	struct rte_mbuf *pkt_segs[RTE_MAX_SEGS_PER_PKT];
166 	struct rte_mbuf *pkt_seg;
167 	uint32_t nb_segs, pkt_len;
168 	uint8_t i;
169 
170 	if (unlikely(tx_pkt_split == TX_PKT_SPLIT_RND))
171 		nb_segs = rte_rand() % tx_pkt_nb_segs + 1;
172 	else
173 		nb_segs = tx_pkt_nb_segs;
174 
175 	if (nb_segs > 1) {
176 		if (rte_mempool_get_bulk(mbp, (void **)pkt_segs, nb_segs - 1))
177 			return false;
178 	}
179 
180 	rte_pktmbuf_reset_headroom(pkt);
181 	pkt->data_len = tx_pkt_seg_lengths[0];
182 	pkt->ol_flags &= EXT_ATTACHED_MBUF;
183 	pkt->ol_flags |= ol_flags;
184 	pkt->vlan_tci = vlan_tci;
185 	pkt->vlan_tci_outer = vlan_tci_outer;
186 	pkt->l2_len = sizeof(struct rte_ether_hdr);
187 	pkt->l3_len = sizeof(struct rte_ipv4_hdr);
188 
189 	pkt_len = pkt->data_len;
190 	pkt_seg = pkt;
191 	for (i = 1; i < nb_segs; i++) {
192 		pkt_seg->next = pkt_segs[i - 1];
193 		pkt_seg = pkt_seg->next;
194 		pkt_seg->data_len = tx_pkt_seg_lengths[i];
195 		pkt_len += pkt_seg->data_len;
196 	}
197 	pkt_seg->next = NULL; /* Last segment of packet. */
198 	/*
199 	 * Copy headers in first packet segment(s).
200 	 */
201 	copy_buf_to_pkt(eth_hdr, sizeof(*eth_hdr), pkt, 0);
202 	copy_buf_to_pkt(&pkt_ip_hdr, sizeof(pkt_ip_hdr), pkt,
203 			sizeof(struct rte_ether_hdr));
204 	if (txonly_multi_flow) {
205 		uint8_t  ip_var = RTE_PER_LCORE(_ip_var);
206 		struct rte_ipv4_hdr *ip_hdr;
207 		uint32_t addr;
208 
209 		ip_hdr = rte_pktmbuf_mtod_offset(pkt,
210 				struct rte_ipv4_hdr *,
211 				sizeof(struct rte_ether_hdr));
212 		/*
213 		 * Generate multiple flows by varying IP src addr. This
214 		 * enables packets are well distributed by RSS in
215 		 * receiver side if any and txonly mode can be a decent
216 		 * packet generator for developer's quick performance
217 		 * regression test.
218 		 */
219 		addr = (tx_ip_dst_addr | (ip_var++ << 8)) + rte_lcore_id();
220 		ip_hdr->src_addr = rte_cpu_to_be_32(addr);
221 		RTE_PER_LCORE(_ip_var) = ip_var;
222 	}
223 	copy_buf_to_pkt(&pkt_udp_hdr, sizeof(pkt_udp_hdr), pkt,
224 			sizeof(struct rte_ether_hdr) +
225 			sizeof(struct rte_ipv4_hdr));
226 	if (unlikely(timestamp_enable)) {
227 		uint64_t skew = RTE_PER_LCORE(timestamp_qskew);
228 		struct {
229 			rte_be32_t signature;
230 			rte_be16_t pkt_idx;
231 			rte_be16_t queue_idx;
232 			rte_be64_t ts;
233 		} timestamp_mark;
234 
235 		if (unlikely(timestamp_init_req !=
236 				RTE_PER_LCORE(timestamp_idone))) {
237 			struct rte_eth_dev *dev = &rte_eth_devices[fs->tx_port];
238 			unsigned int txqs_n = dev->data->nb_tx_queues;
239 			uint64_t phase = tx_pkt_times_inter * fs->tx_queue /
240 					 (txqs_n ? txqs_n : 1);
241 			/*
242 			 * Initialize the scheduling time phase shift
243 			 * depending on queue index.
244 			 */
245 			skew = timestamp_initial[fs->tx_port] +
246 			       tx_pkt_times_inter + phase;
247 			RTE_PER_LCORE(timestamp_qskew) = skew;
248 			RTE_PER_LCORE(timestamp_idone) = timestamp_init_req;
249 		}
250 		timestamp_mark.pkt_idx = rte_cpu_to_be_16(idx);
251 		timestamp_mark.queue_idx = rte_cpu_to_be_16(fs->tx_queue);
252 		timestamp_mark.signature = rte_cpu_to_be_32(0xBEEFC0DE);
253 		if (unlikely(!idx)) {
254 			skew +=	tx_pkt_times_inter;
255 			pkt->ol_flags |= timestamp_mask;
256 			*RTE_MBUF_DYNFIELD
257 				(pkt, timestamp_off, uint64_t *) = skew;
258 			RTE_PER_LCORE(timestamp_qskew) = skew;
259 			timestamp_mark.ts = rte_cpu_to_be_64(skew);
260 		} else if (tx_pkt_times_intra) {
261 			skew +=	tx_pkt_times_intra;
262 			pkt->ol_flags |= timestamp_mask;
263 			*RTE_MBUF_DYNFIELD
264 				(pkt, timestamp_off, uint64_t *) = skew;
265 			RTE_PER_LCORE(timestamp_qskew) = skew;
266 			timestamp_mark.ts = rte_cpu_to_be_64(skew);
267 		} else {
268 			timestamp_mark.ts = RTE_BE64(0);
269 		}
270 		copy_buf_to_pkt(&timestamp_mark, sizeof(timestamp_mark), pkt,
271 			sizeof(struct rte_ether_hdr) +
272 			sizeof(struct rte_ipv4_hdr) +
273 			sizeof(pkt_udp_hdr));
274 	}
275 	/*
276 	 * Complete first mbuf of packet and append it to the
277 	 * burst of packets to be transmitted.
278 	 */
279 	pkt->nb_segs = nb_segs;
280 	pkt->pkt_len = pkt_len;
281 
282 	return true;
283 }
284 
285 /*
286  * Transmit a burst of multi-segments packets.
287  */
288 static void
289 pkt_burst_transmit(struct fwd_stream *fs)
290 {
291 	struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
292 	struct rte_port *txp;
293 	struct rte_mbuf *pkt;
294 	struct rte_mempool *mbp;
295 	struct rte_ether_hdr eth_hdr;
296 	uint16_t nb_tx;
297 	uint16_t nb_pkt;
298 	uint16_t vlan_tci, vlan_tci_outer;
299 	uint32_t retry;
300 	uint64_t ol_flags = 0;
301 	uint64_t tx_offloads;
302 	uint64_t start_tsc = 0;
303 
304 	get_start_cycles(&start_tsc);
305 
306 	mbp = current_fwd_lcore()->mbp;
307 	txp = &ports[fs->tx_port];
308 	tx_offloads = txp->dev_conf.txmode.offloads;
309 	vlan_tci = txp->tx_vlan_id;
310 	vlan_tci_outer = txp->tx_vlan_id_outer;
311 	if (tx_offloads	& DEV_TX_OFFLOAD_VLAN_INSERT)
312 		ol_flags = PKT_TX_VLAN_PKT;
313 	if (tx_offloads & DEV_TX_OFFLOAD_QINQ_INSERT)
314 		ol_flags |= PKT_TX_QINQ_PKT;
315 	if (tx_offloads & DEV_TX_OFFLOAD_MACSEC_INSERT)
316 		ol_flags |= PKT_TX_MACSEC;
317 
318 	/*
319 	 * Initialize Ethernet header.
320 	 */
321 	rte_ether_addr_copy(&peer_eth_addrs[fs->peer_addr], &eth_hdr.d_addr);
322 	rte_ether_addr_copy(&ports[fs->tx_port].eth_addr, &eth_hdr.s_addr);
323 	eth_hdr.ether_type = rte_cpu_to_be_16(RTE_ETHER_TYPE_IPV4);
324 
325 	if (rte_mempool_get_bulk(mbp, (void **)pkts_burst,
326 				nb_pkt_per_burst) == 0) {
327 		for (nb_pkt = 0; nb_pkt < nb_pkt_per_burst; nb_pkt++) {
328 			if (unlikely(!pkt_burst_prepare(pkts_burst[nb_pkt], mbp,
329 							&eth_hdr, vlan_tci,
330 							vlan_tci_outer,
331 							ol_flags,
332 							nb_pkt, fs))) {
333 				rte_mempool_put_bulk(mbp,
334 						(void **)&pkts_burst[nb_pkt],
335 						nb_pkt_per_burst - nb_pkt);
336 				break;
337 			}
338 		}
339 	} else {
340 		for (nb_pkt = 0; nb_pkt < nb_pkt_per_burst; nb_pkt++) {
341 			pkt = rte_mbuf_raw_alloc(mbp);
342 			if (pkt == NULL)
343 				break;
344 			if (unlikely(!pkt_burst_prepare(pkt, mbp, &eth_hdr,
345 							vlan_tci,
346 							vlan_tci_outer,
347 							ol_flags,
348 							nb_pkt, fs))) {
349 				rte_pktmbuf_free(pkt);
350 				break;
351 			}
352 			pkts_burst[nb_pkt] = pkt;
353 		}
354 	}
355 
356 	if (nb_pkt == 0)
357 		return;
358 
359 	nb_tx = rte_eth_tx_burst(fs->tx_port, fs->tx_queue, pkts_burst, nb_pkt);
360 
361 	/*
362 	 * Retry if necessary
363 	 */
364 	if (unlikely(nb_tx < nb_pkt) && fs->retry_enabled) {
365 		retry = 0;
366 		while (nb_tx < nb_pkt && retry++ < burst_tx_retry_num) {
367 			rte_delay_us(burst_tx_delay_time);
368 			nb_tx += rte_eth_tx_burst(fs->tx_port, fs->tx_queue,
369 					&pkts_burst[nb_tx], nb_pkt - nb_tx);
370 		}
371 	}
372 	fs->tx_packets += nb_tx;
373 
374 	if (txonly_multi_flow)
375 		RTE_PER_LCORE(_ip_var) -= nb_pkt - nb_tx;
376 
377 	inc_tx_burst_stats(fs, nb_tx);
378 	if (unlikely(nb_tx < nb_pkt)) {
379 		if (verbose_level > 0 && fs->fwd_dropped == 0)
380 			printf("port %d tx_queue %d - drop "
381 			       "(nb_pkt:%u - nb_tx:%u)=%u packets\n",
382 			       fs->tx_port, fs->tx_queue,
383 			       (unsigned) nb_pkt, (unsigned) nb_tx,
384 			       (unsigned) (nb_pkt - nb_tx));
385 		fs->fwd_dropped += (nb_pkt - nb_tx);
386 		do {
387 			rte_pktmbuf_free(pkts_burst[nb_tx]);
388 		} while (++nb_tx < nb_pkt);
389 	}
390 
391 	get_end_cycles(fs, start_tsc);
392 }
393 
394 static void
395 tx_only_begin(portid_t pi)
396 {
397 	uint16_t pkt_data_len;
398 	int dynf;
399 
400 	pkt_data_len = (uint16_t) (tx_pkt_length - (
401 					sizeof(struct rte_ether_hdr) +
402 					sizeof(struct rte_ipv4_hdr) +
403 					sizeof(struct rte_udp_hdr)));
404 	setup_pkt_udp_ip_headers(&pkt_ip_hdr, &pkt_udp_hdr, pkt_data_len);
405 
406 	timestamp_enable = false;
407 	timestamp_mask = 0;
408 	timestamp_off = -1;
409 	RTE_PER_LCORE(timestamp_qskew) = 0;
410 	dynf = rte_mbuf_dynflag_lookup
411 				(RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME, NULL);
412 	if (dynf >= 0)
413 		timestamp_mask = 1ULL << dynf;
414 	dynf = rte_mbuf_dynfield_lookup
415 				(RTE_MBUF_DYNFIELD_TIMESTAMP_NAME, NULL);
416 	if (dynf >= 0)
417 		timestamp_off = dynf;
418 	timestamp_enable = tx_pkt_times_inter &&
419 			   timestamp_mask &&
420 			   timestamp_off >= 0 &&
421 			   !rte_eth_read_clock(pi, &timestamp_initial[pi]);
422 	if (timestamp_enable)
423 		timestamp_init_req++;
424 	/* Make sure all settings are visible on forwarding cores.*/
425 	rte_wmb();
426 }
427 
428 struct fwd_engine tx_only_engine = {
429 	.fwd_mode_name  = "txonly",
430 	.port_fwd_begin = tx_only_begin,
431 	.port_fwd_end   = NULL,
432 	.packet_fwd     = pkt_burst_transmit,
433 };
434