1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
13 #include <sys/queue.h>
16 #include <rte_common.h>
17 #include <rte_byteorder.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>
25 #include <rte_per_lcore.h>
26 #include <rte_lcore.h>
27 #include <rte_branch_prediction.h>
28 #include <rte_mempool.h>
30 #include <rte_interrupts.h>
32 #include <rte_ether.h>
33 #include <rte_ethdev.h>
37 #include <rte_string_fns.h>
49 /* use RFC863 Discard Protocol */
50 uint16_t tx_udp_src_port = 9;
51 uint16_t tx_udp_dst_port = 9;
53 /* use RFC5735 / RFC2544 reserved network test addresses */
54 uint32_t tx_ip_src_addr = (198U << 24) | (18 << 16) | (0 << 8) | 1;
55 uint32_t tx_ip_dst_addr = (198U << 24) | (18 << 16) | (0 << 8) | 2;
57 #define IP_DEFTTL 64 /* from RFC 1340. */
59 static struct rte_ipv4_hdr pkt_ip_hdr; /**< IP header of transmitted packets. */
60 RTE_DEFINE_PER_LCORE(uint8_t, _ip_var); /**< IP address variation */
61 static struct rte_udp_hdr pkt_udp_hdr; /**< UDP header of tx packets. */
63 static uint64_t timestamp_mask; /**< Timestamp dynamic flag mask */
64 static int32_t timestamp_off; /**< Timestamp dynamic field offset */
65 static bool timestamp_enable; /**< Timestamp enable */
66 static uint64_t timestamp_initial[RTE_MAX_ETHPORTS];
69 copy_buf_to_pkt_segs(void* buf, unsigned len, struct rte_mbuf *pkt,
77 while (offset >= seg->data_len) {
78 offset -= seg->data_len;
81 copy_len = seg->data_len - offset;
82 seg_buf = rte_pktmbuf_mtod_offset(seg, char *, offset);
83 while (len > copy_len) {
84 rte_memcpy(seg_buf, buf, (size_t) copy_len);
86 buf = ((char*) buf + copy_len);
88 seg_buf = rte_pktmbuf_mtod(seg, char *);
89 copy_len = seg->data_len;
91 rte_memcpy(seg_buf, buf, (size_t) len);
95 copy_buf_to_pkt(void* buf, unsigned len, struct rte_mbuf *pkt, unsigned offset)
97 if (offset + len <= pkt->data_len) {
98 rte_memcpy(rte_pktmbuf_mtod_offset(pkt, char *, offset),
102 copy_buf_to_pkt_segs(buf, len, pkt, offset);
106 setup_pkt_udp_ip_headers(struct rte_ipv4_hdr *ip_hdr,
107 struct rte_udp_hdr *udp_hdr,
108 uint16_t pkt_data_len)
115 * Initialize UDP header.
117 pkt_len = (uint16_t) (pkt_data_len + sizeof(struct rte_udp_hdr));
118 udp_hdr->src_port = rte_cpu_to_be_16(tx_udp_src_port);
119 udp_hdr->dst_port = rte_cpu_to_be_16(tx_udp_dst_port);
120 udp_hdr->dgram_len = RTE_CPU_TO_BE_16(pkt_len);
121 udp_hdr->dgram_cksum = 0; /* No UDP checksum. */
124 * Initialize IP header.
126 pkt_len = (uint16_t) (pkt_len + sizeof(struct rte_ipv4_hdr));
127 ip_hdr->version_ihl = RTE_IPV4_VHL_DEF;
128 ip_hdr->type_of_service = 0;
129 ip_hdr->fragment_offset = 0;
130 ip_hdr->time_to_live = IP_DEFTTL;
131 ip_hdr->next_proto_id = IPPROTO_UDP;
132 ip_hdr->packet_id = 0;
133 ip_hdr->total_length = RTE_CPU_TO_BE_16(pkt_len);
134 ip_hdr->src_addr = rte_cpu_to_be_32(tx_ip_src_addr);
135 ip_hdr->dst_addr = rte_cpu_to_be_32(tx_ip_dst_addr);
138 * Compute IP header checksum.
140 ptr16 = (unaligned_uint16_t*) ip_hdr;
142 ip_cksum += ptr16[0]; ip_cksum += ptr16[1];
143 ip_cksum += ptr16[2]; ip_cksum += ptr16[3];
144 ip_cksum += ptr16[4];
145 ip_cksum += ptr16[6]; ip_cksum += ptr16[7];
146 ip_cksum += ptr16[8]; ip_cksum += ptr16[9];
149 * Reduce 32 bit checksum to 16 bits and complement it.
151 ip_cksum = ((ip_cksum & 0xFFFF0000) >> 16) +
152 (ip_cksum & 0x0000FFFF);
153 if (ip_cksum > 65535)
155 ip_cksum = (~ip_cksum) & 0x0000FFFF;
158 ip_hdr->hdr_checksum = (uint16_t) ip_cksum;
162 update_pkt_header(struct rte_mbuf *pkt, uint32_t total_pkt_len)
164 struct rte_ipv4_hdr *ip_hdr;
165 struct rte_udp_hdr *udp_hdr;
166 uint16_t pkt_data_len;
169 pkt_data_len = (uint16_t) (total_pkt_len - (
170 sizeof(struct rte_ether_hdr) +
171 sizeof(struct rte_ipv4_hdr) +
172 sizeof(struct rte_udp_hdr)));
173 /* update UDP packet length */
174 udp_hdr = rte_pktmbuf_mtod_offset(pkt, struct rte_udp_hdr *,
175 sizeof(struct rte_ether_hdr) +
176 sizeof(struct rte_ipv4_hdr));
177 pkt_len = (uint16_t) (pkt_data_len + sizeof(struct rte_udp_hdr));
178 udp_hdr->dgram_len = RTE_CPU_TO_BE_16(pkt_len);
180 /* update IP packet length and checksum */
181 ip_hdr = rte_pktmbuf_mtod_offset(pkt, struct rte_ipv4_hdr *,
182 sizeof(struct rte_ether_hdr));
183 ip_hdr->hdr_checksum = 0;
184 pkt_len = (uint16_t) (pkt_len + sizeof(struct rte_ipv4_hdr));
185 ip_hdr->total_length = RTE_CPU_TO_BE_16(pkt_len);
186 ip_hdr->hdr_checksum = rte_ipv4_cksum(ip_hdr);
190 pkt_burst_prepare(struct rte_mbuf *pkt, struct rte_mempool *mbp,
191 struct rte_ether_hdr *eth_hdr, const uint16_t vlan_tci,
192 const uint16_t vlan_tci_outer, const uint64_t ol_flags,
193 const uint16_t idx, struct fwd_stream *fs)
195 struct rte_mbuf *pkt_segs[RTE_MAX_SEGS_PER_PKT];
196 struct rte_mbuf *pkt_seg;
197 uint32_t nb_segs, pkt_len;
200 if (unlikely(tx_pkt_split == TX_PKT_SPLIT_RND))
201 nb_segs = rte_rand() % tx_pkt_nb_segs + 1;
203 nb_segs = tx_pkt_nb_segs;
206 if (rte_mempool_get_bulk(mbp, (void **)pkt_segs, nb_segs - 1))
210 rte_pktmbuf_reset_headroom(pkt);
211 pkt->data_len = tx_pkt_seg_lengths[0];
212 pkt->ol_flags &= RTE_MBUF_F_EXTERNAL;
213 pkt->ol_flags |= ol_flags;
214 pkt->vlan_tci = vlan_tci;
215 pkt->vlan_tci_outer = vlan_tci_outer;
216 pkt->l2_len = sizeof(struct rte_ether_hdr);
217 pkt->l3_len = sizeof(struct rte_ipv4_hdr);
219 pkt_len = pkt->data_len;
221 for (i = 1; i < nb_segs; i++) {
222 pkt_seg->next = pkt_segs[i - 1];
223 pkt_seg = pkt_seg->next;
224 pkt_seg->data_len = tx_pkt_seg_lengths[i];
225 pkt_len += pkt_seg->data_len;
227 pkt_seg->next = NULL; /* Last segment of packet. */
229 * Copy headers in first packet segment(s).
231 copy_buf_to_pkt(eth_hdr, sizeof(*eth_hdr), pkt, 0);
232 copy_buf_to_pkt(&pkt_ip_hdr, sizeof(pkt_ip_hdr), pkt,
233 sizeof(struct rte_ether_hdr));
234 if (txonly_multi_flow) {
235 uint8_t ip_var = RTE_PER_LCORE(_ip_var);
236 struct rte_ipv4_hdr *ip_hdr;
239 ip_hdr = rte_pktmbuf_mtod_offset(pkt,
240 struct rte_ipv4_hdr *,
241 sizeof(struct rte_ether_hdr));
243 * Generate multiple flows by varying IP src addr. This
244 * enables packets are well distributed by RSS in
245 * receiver side if any and txonly mode can be a decent
246 * packet generator for developer's quick performance
249 addr = (tx_ip_dst_addr | (ip_var++ << 8)) + rte_lcore_id();
250 ip_hdr->src_addr = rte_cpu_to_be_32(addr);
251 RTE_PER_LCORE(_ip_var) = ip_var;
253 copy_buf_to_pkt(&pkt_udp_hdr, sizeof(pkt_udp_hdr), pkt,
254 sizeof(struct rte_ether_hdr) +
255 sizeof(struct rte_ipv4_hdr));
257 if (unlikely(tx_pkt_split == TX_PKT_SPLIT_RND) || txonly_multi_flow)
258 update_pkt_header(pkt, pkt_len);
260 if (unlikely(timestamp_enable)) {
261 uint64_t skew = fs->ts_skew;
262 struct tx_timestamp timestamp_mark;
264 if (unlikely(!skew)) {
265 struct rte_eth_dev_info dev_info;
270 ret = eth_dev_info_get_print_err(fs->tx_port, &dev_info);
273 "Failed to get device info for port %d,"
274 "could not finish timestamp init",
278 txqs_n = dev_info.nb_tx_queues;
279 phase = tx_pkt_times_inter * fs->tx_queue /
280 (txqs_n ? txqs_n : 1);
282 * Initialize the scheduling time phase shift
283 * depending on queue index.
285 skew = timestamp_initial[fs->tx_port] +
286 tx_pkt_times_inter + phase;
289 timestamp_mark.pkt_idx = rte_cpu_to_be_16(idx);
290 timestamp_mark.queue_idx = rte_cpu_to_be_16(fs->tx_queue);
291 timestamp_mark.signature = rte_cpu_to_be_32(0xBEEFC0DE);
292 if (unlikely(!idx)) {
293 skew += tx_pkt_times_inter;
294 pkt->ol_flags |= timestamp_mask;
296 (pkt, timestamp_off, uint64_t *) = skew;
298 timestamp_mark.ts = rte_cpu_to_be_64(skew);
299 } else if (tx_pkt_times_intra) {
300 skew += tx_pkt_times_intra;
301 pkt->ol_flags |= timestamp_mask;
303 (pkt, timestamp_off, uint64_t *) = skew;
305 timestamp_mark.ts = rte_cpu_to_be_64(skew);
307 timestamp_mark.ts = RTE_BE64(0);
309 copy_buf_to_pkt(×tamp_mark, sizeof(timestamp_mark), pkt,
310 sizeof(struct rte_ether_hdr) +
311 sizeof(struct rte_ipv4_hdr) +
312 sizeof(pkt_udp_hdr));
315 * Complete first mbuf of packet and append it to the
316 * burst of packets to be transmitted.
318 pkt->nb_segs = nb_segs;
319 pkt->pkt_len = pkt_len;
325 * Transmit a burst of multi-segments packets.
328 pkt_burst_transmit(struct fwd_stream *fs)
330 struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
331 struct rte_port *txp;
332 struct rte_mbuf *pkt;
333 struct rte_mempool *mbp;
334 struct rte_ether_hdr eth_hdr;
337 uint16_t vlan_tci, vlan_tci_outer;
339 uint64_t ol_flags = 0;
340 uint64_t tx_offloads;
341 uint64_t start_tsc = 0;
343 get_start_cycles(&start_tsc);
345 mbp = current_fwd_lcore()->mbp;
346 txp = &ports[fs->tx_port];
347 tx_offloads = txp->dev_conf.txmode.offloads;
348 vlan_tci = txp->tx_vlan_id;
349 vlan_tci_outer = txp->tx_vlan_id_outer;
350 if (tx_offloads & RTE_ETH_TX_OFFLOAD_VLAN_INSERT)
351 ol_flags = RTE_MBUF_F_TX_VLAN;
352 if (tx_offloads & RTE_ETH_TX_OFFLOAD_QINQ_INSERT)
353 ol_flags |= RTE_MBUF_F_TX_QINQ;
354 if (tx_offloads & RTE_ETH_TX_OFFLOAD_MACSEC_INSERT)
355 ol_flags |= RTE_MBUF_F_TX_MACSEC;
358 * Initialize Ethernet header.
360 rte_ether_addr_copy(&peer_eth_addrs[fs->peer_addr], ð_hdr.dst_addr);
361 rte_ether_addr_copy(&ports[fs->tx_port].eth_addr, ð_hdr.src_addr);
362 eth_hdr.ether_type = rte_cpu_to_be_16(RTE_ETHER_TYPE_IPV4);
364 if (rte_mempool_get_bulk(mbp, (void **)pkts_burst,
365 nb_pkt_per_burst) == 0) {
366 for (nb_pkt = 0; nb_pkt < nb_pkt_per_burst; nb_pkt++) {
367 if (unlikely(!pkt_burst_prepare(pkts_burst[nb_pkt], mbp,
372 rte_mempool_put_bulk(mbp,
373 (void **)&pkts_burst[nb_pkt],
374 nb_pkt_per_burst - nb_pkt);
379 for (nb_pkt = 0; nb_pkt < nb_pkt_per_burst; nb_pkt++) {
380 pkt = rte_mbuf_raw_alloc(mbp);
383 if (unlikely(!pkt_burst_prepare(pkt, mbp, ð_hdr,
388 rte_pktmbuf_free(pkt);
391 pkts_burst[nb_pkt] = pkt;
398 nb_tx = rte_eth_tx_burst(fs->tx_port, fs->tx_queue, pkts_burst, nb_pkt);
403 if (unlikely(nb_tx < nb_pkt) && fs->retry_enabled) {
405 while (nb_tx < nb_pkt && retry++ < burst_tx_retry_num) {
406 rte_delay_us(burst_tx_delay_time);
407 nb_tx += rte_eth_tx_burst(fs->tx_port, fs->tx_queue,
408 &pkts_burst[nb_tx], nb_pkt - nb_tx);
411 fs->tx_packets += nb_tx;
413 if (txonly_multi_flow)
414 RTE_PER_LCORE(_ip_var) -= nb_pkt - nb_tx;
416 inc_tx_burst_stats(fs, nb_tx);
417 if (unlikely(nb_tx < nb_pkt)) {
418 if (verbose_level > 0 && fs->fwd_dropped == 0)
419 printf("port %d tx_queue %d - drop "
420 "(nb_pkt:%u - nb_tx:%u)=%u packets\n",
421 fs->tx_port, fs->tx_queue,
422 (unsigned) nb_pkt, (unsigned) nb_tx,
423 (unsigned) (nb_pkt - nb_tx));
424 fs->fwd_dropped += (nb_pkt - nb_tx);
426 rte_pktmbuf_free(pkts_burst[nb_tx]);
427 } while (++nb_tx < nb_pkt);
430 get_end_cycles(fs, start_tsc);
434 tx_only_begin(portid_t pi)
436 uint16_t pkt_hdr_len, pkt_data_len;
439 pkt_hdr_len = (uint16_t)(sizeof(struct rte_ether_hdr) +
440 sizeof(struct rte_ipv4_hdr) +
441 sizeof(struct rte_udp_hdr));
442 pkt_data_len = tx_pkt_length - pkt_hdr_len;
444 if ((tx_pkt_split == TX_PKT_SPLIT_RND || txonly_multi_flow) &&
445 tx_pkt_seg_lengths[0] < pkt_hdr_len) {
447 "Random segment number or multiple flow is enabled, "
448 "but tx_pkt_seg_lengths[0] %u < %u (needed)\n",
449 tx_pkt_seg_lengths[0], pkt_hdr_len);
453 setup_pkt_udp_ip_headers(&pkt_ip_hdr, &pkt_udp_hdr, pkt_data_len);
455 timestamp_enable = false;
458 dynf = rte_mbuf_dynflag_lookup
459 (RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME, NULL);
461 timestamp_mask = 1ULL << dynf;
462 dynf = rte_mbuf_dynfield_lookup
463 (RTE_MBUF_DYNFIELD_TIMESTAMP_NAME, NULL);
465 timestamp_off = dynf;
466 timestamp_enable = tx_pkt_times_inter &&
468 timestamp_off >= 0 &&
469 !rte_eth_read_clock(pi, ×tamp_initial[pi]);
471 if (timestamp_enable) {
472 pkt_hdr_len += sizeof(struct tx_timestamp);
474 if (tx_pkt_split == TX_PKT_SPLIT_RND) {
475 if (tx_pkt_seg_lengths[0] < pkt_hdr_len) {
477 "Time stamp and random segment number are enabled, "
478 "but tx_pkt_seg_lengths[0] %u < %u (needed)\n",
479 tx_pkt_seg_lengths[0], pkt_hdr_len);
486 for (i = 0; i < tx_pkt_nb_segs; i++) {
487 total += tx_pkt_seg_lengths[i];
488 if (total >= pkt_hdr_len)
492 if (total < pkt_hdr_len) {
494 "Not enough Tx segment space for time stamp info, "
495 "total %u < %u (needed)\n",
502 /* Make sure all settings are visible on forwarding cores.*/
507 struct fwd_engine tx_only_engine = {
508 .fwd_mode_name = "txonly",
509 .port_fwd_begin = tx_only_begin,
510 .port_fwd_end = NULL,
511 .packet_fwd = pkt_burst_transmit,