1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2010-2015 Intel Corporation 3 */ 4 5 #include <stdint.h> 6 #include <inttypes.h> 7 #include <getopt.h> 8 #include <rte_eal.h> 9 #include <rte_ethdev.h> 10 #include <rte_cycles.h> 11 #include <rte_lcore.h> 12 #include <rte_mbuf.h> 13 #include <rte_mbuf_dyn.h> 14 15 #define RX_RING_SIZE 1024 16 #define TX_RING_SIZE 1024 17 18 #define NUM_MBUFS 8191 19 #define MBUF_CACHE_SIZE 250 20 #define BURST_SIZE 32 21 22 static int hwts_dynfield_offset = -1; 23 24 static inline rte_mbuf_timestamp_t * 25 hwts_field(struct rte_mbuf *mbuf) 26 { 27 return RTE_MBUF_DYNFIELD(mbuf, 28 hwts_dynfield_offset, rte_mbuf_timestamp_t *); 29 } 30 31 typedef uint64_t tsc_t; 32 static int tsc_dynfield_offset = -1; 33 34 static inline tsc_t * 35 tsc_field(struct rte_mbuf *mbuf) 36 { 37 return RTE_MBUF_DYNFIELD(mbuf, tsc_dynfield_offset, tsc_t *); 38 } 39 40 static const char usage[] = 41 "%s EAL_ARGS -- [-t]\n"; 42 43 static const struct rte_eth_conf port_conf_default = { 44 .rxmode = { 45 .max_rx_pkt_len = RTE_ETHER_MAX_LEN, 46 }, 47 }; 48 49 static struct { 50 uint64_t total_cycles; 51 uint64_t total_queue_cycles; 52 uint64_t total_pkts; 53 } latency_numbers; 54 55 int hw_timestamping; 56 57 #define TICKS_PER_CYCLE_SHIFT 16 58 static uint64_t ticks_per_cycle_mult; 59 60 /* Callback added to the RX port and applied to packets. 8< */ 61 static uint16_t 62 add_timestamps(uint16_t port __rte_unused, uint16_t qidx __rte_unused, 63 struct rte_mbuf **pkts, uint16_t nb_pkts, 64 uint16_t max_pkts __rte_unused, void *_ __rte_unused) 65 { 66 unsigned i; 67 uint64_t now = rte_rdtsc(); 68 69 for (i = 0; i < nb_pkts; i++) 70 *tsc_field(pkts[i]) = now; 71 return nb_pkts; 72 } 73 /* >8 End of callback addition and application. */ 74 75 /* Callback is added to the TX port. 8< */ 76 static uint16_t 77 calc_latency(uint16_t port, uint16_t qidx __rte_unused, 78 struct rte_mbuf **pkts, uint16_t nb_pkts, void *_ __rte_unused) 79 { 80 uint64_t cycles = 0; 81 uint64_t queue_ticks = 0; 82 uint64_t now = rte_rdtsc(); 83 uint64_t ticks; 84 unsigned i; 85 86 if (hw_timestamping) 87 rte_eth_read_clock(port, &ticks); 88 89 for (i = 0; i < nb_pkts; i++) { 90 cycles += now - *tsc_field(pkts[i]); 91 if (hw_timestamping) 92 queue_ticks += ticks - *hwts_field(pkts[i]); 93 } 94 95 latency_numbers.total_cycles += cycles; 96 if (hw_timestamping) 97 latency_numbers.total_queue_cycles += (queue_ticks 98 * ticks_per_cycle_mult) >> TICKS_PER_CYCLE_SHIFT; 99 100 latency_numbers.total_pkts += nb_pkts; 101 102 if (latency_numbers.total_pkts > (100 * 1000 * 1000ULL)) { 103 printf("Latency = %"PRIu64" cycles\n", 104 latency_numbers.total_cycles / latency_numbers.total_pkts); 105 if (hw_timestamping) { 106 printf("Latency from HW = %"PRIu64" cycles\n", 107 latency_numbers.total_queue_cycles 108 / latency_numbers.total_pkts); 109 } 110 latency_numbers.total_cycles = 0; 111 latency_numbers.total_queue_cycles = 0; 112 latency_numbers.total_pkts = 0; 113 } 114 return nb_pkts; 115 } 116 /* >8 End of callback addition. */ 117 118 /* 119 * Initialises a given port using global settings and with the rx buffers 120 * coming from the mbuf_pool passed as parameter 121 */ 122 123 /* Port initialization. 8< */ 124 static inline int 125 port_init(uint16_t port, struct rte_mempool *mbuf_pool) 126 { 127 struct rte_eth_conf port_conf = port_conf_default; 128 const uint16_t rx_rings = 1, tx_rings = 1; 129 uint16_t nb_rxd = RX_RING_SIZE; 130 uint16_t nb_txd = TX_RING_SIZE; 131 int retval; 132 uint16_t q; 133 struct rte_eth_dev_info dev_info; 134 struct rte_eth_rxconf rxconf; 135 struct rte_eth_txconf txconf; 136 137 if (!rte_eth_dev_is_valid_port(port)) 138 return -1; 139 140 retval = rte_eth_dev_info_get(port, &dev_info); 141 if (retval != 0) { 142 printf("Error during getting device (port %u) info: %s\n", 143 port, strerror(-retval)); 144 145 return retval; 146 } 147 148 if (dev_info.tx_offload_capa & DEV_TX_OFFLOAD_MBUF_FAST_FREE) 149 port_conf.txmode.offloads |= 150 DEV_TX_OFFLOAD_MBUF_FAST_FREE; 151 152 if (hw_timestamping) { 153 if (!(dev_info.rx_offload_capa & DEV_RX_OFFLOAD_TIMESTAMP)) { 154 printf("\nERROR: Port %u does not support hardware timestamping\n" 155 , port); 156 return -1; 157 } 158 port_conf.rxmode.offloads |= DEV_RX_OFFLOAD_TIMESTAMP; 159 rte_mbuf_dyn_rx_timestamp_register(&hwts_dynfield_offset, NULL); 160 if (hwts_dynfield_offset < 0) { 161 printf("ERROR: Failed to register timestamp field\n"); 162 return -rte_errno; 163 } 164 } 165 166 retval = rte_eth_dev_configure(port, rx_rings, tx_rings, &port_conf); 167 if (retval != 0) 168 return retval; 169 170 retval = rte_eth_dev_adjust_nb_rx_tx_desc(port, &nb_rxd, &nb_txd); 171 if (retval != 0) 172 return retval; 173 174 rxconf = dev_info.default_rxconf; 175 176 for (q = 0; q < rx_rings; q++) { 177 retval = rte_eth_rx_queue_setup(port, q, nb_rxd, 178 rte_eth_dev_socket_id(port), &rxconf, mbuf_pool); 179 if (retval < 0) 180 return retval; 181 } 182 183 txconf = dev_info.default_txconf; 184 txconf.offloads = port_conf.txmode.offloads; 185 for (q = 0; q < tx_rings; q++) { 186 retval = rte_eth_tx_queue_setup(port, q, nb_txd, 187 rte_eth_dev_socket_id(port), &txconf); 188 if (retval < 0) 189 return retval; 190 } 191 192 retval = rte_eth_dev_start(port); 193 if (retval < 0) 194 return retval; 195 196 if (hw_timestamping && ticks_per_cycle_mult == 0) { 197 uint64_t cycles_base = rte_rdtsc(); 198 uint64_t ticks_base; 199 retval = rte_eth_read_clock(port, &ticks_base); 200 if (retval != 0) 201 return retval; 202 rte_delay_ms(100); 203 uint64_t cycles = rte_rdtsc(); 204 uint64_t ticks; 205 rte_eth_read_clock(port, &ticks); 206 uint64_t c_freq = cycles - cycles_base; 207 uint64_t t_freq = ticks - ticks_base; 208 double freq_mult = (double)c_freq / t_freq; 209 printf("TSC Freq ~= %" PRIu64 210 "\nHW Freq ~= %" PRIu64 211 "\nRatio : %f\n", 212 c_freq * 10, t_freq * 10, freq_mult); 213 /* TSC will be faster than internal ticks so freq_mult is > 0 214 * We convert the multiplication to an integer shift & mult 215 */ 216 ticks_per_cycle_mult = (1 << TICKS_PER_CYCLE_SHIFT) / freq_mult; 217 } 218 219 struct rte_ether_addr addr; 220 221 retval = rte_eth_macaddr_get(port, &addr); 222 if (retval < 0) { 223 printf("Failed to get MAC address on port %u: %s\n", 224 port, rte_strerror(-retval)); 225 return retval; 226 } 227 printf("Port %u MAC: %02"PRIx8" %02"PRIx8" %02"PRIx8 228 " %02"PRIx8" %02"PRIx8" %02"PRIx8"\n", 229 (unsigned)port, 230 addr.addr_bytes[0], addr.addr_bytes[1], 231 addr.addr_bytes[2], addr.addr_bytes[3], 232 addr.addr_bytes[4], addr.addr_bytes[5]); 233 234 retval = rte_eth_promiscuous_enable(port); 235 if (retval != 0) 236 return retval; 237 238 /* RX and TX callbacks are added to the ports. 8< */ 239 rte_eth_add_rx_callback(port, 0, add_timestamps, NULL); 240 rte_eth_add_tx_callback(port, 0, calc_latency, NULL); 241 /* >8 End of RX and TX callbacks. */ 242 243 return 0; 244 } 245 /* >8 End of port initialization. */ 246 247 /* 248 * Main thread that does the work, reading from INPUT_PORT 249 * and writing to OUTPUT_PORT 250 */ 251 static __rte_noreturn void 252 lcore_main(void) 253 { 254 uint16_t port; 255 256 RTE_ETH_FOREACH_DEV(port) 257 if (rte_eth_dev_socket_id(port) > 0 && 258 rte_eth_dev_socket_id(port) != 259 (int)rte_socket_id()) 260 printf("WARNING, port %u is on remote NUMA node to " 261 "polling thread.\n\tPerformance will " 262 "not be optimal.\n", port); 263 264 printf("\nCore %u forwarding packets. [Ctrl+C to quit]\n", 265 rte_lcore_id()); 266 for (;;) { 267 RTE_ETH_FOREACH_DEV(port) { 268 struct rte_mbuf *bufs[BURST_SIZE]; 269 const uint16_t nb_rx = rte_eth_rx_burst(port, 0, 270 bufs, BURST_SIZE); 271 if (unlikely(nb_rx == 0)) 272 continue; 273 const uint16_t nb_tx = rte_eth_tx_burst(port ^ 1, 0, 274 bufs, nb_rx); 275 if (unlikely(nb_tx < nb_rx)) { 276 uint16_t buf; 277 278 for (buf = nb_tx; buf < nb_rx; buf++) 279 rte_pktmbuf_free(bufs[buf]); 280 } 281 } 282 } 283 } 284 285 /* Main function, does initialisation and calls the per-lcore functions */ 286 int 287 main(int argc, char *argv[]) 288 { 289 struct rte_mempool *mbuf_pool; 290 uint16_t nb_ports; 291 uint16_t portid; 292 struct option lgopts[] = { 293 { NULL, 0, 0, 0 } 294 }; 295 int opt, option_index; 296 297 static const struct rte_mbuf_dynfield tsc_dynfield_desc = { 298 .name = "example_bbdev_dynfield_tsc", 299 .size = sizeof(tsc_t), 300 .align = __alignof__(tsc_t), 301 }; 302 303 /* init EAL */ 304 int ret = rte_eal_init(argc, argv); 305 306 if (ret < 0) 307 rte_exit(EXIT_FAILURE, "Error with EAL initialization\n"); 308 argc -= ret; 309 argv += ret; 310 311 while ((opt = getopt_long(argc, argv, "t", lgopts, &option_index)) 312 != EOF) 313 switch (opt) { 314 case 't': 315 hw_timestamping = 1; 316 break; 317 default: 318 printf(usage, argv[0]); 319 return -1; 320 } 321 optind = 1; /* reset getopt lib */ 322 323 nb_ports = rte_eth_dev_count_avail(); 324 if (nb_ports < 2 || (nb_ports & 1)) 325 rte_exit(EXIT_FAILURE, "Error: number of ports must be even\n"); 326 327 mbuf_pool = rte_pktmbuf_pool_create("MBUF_POOL", 328 NUM_MBUFS * nb_ports, MBUF_CACHE_SIZE, 0, 329 RTE_MBUF_DEFAULT_BUF_SIZE, rte_socket_id()); 330 if (mbuf_pool == NULL) 331 rte_exit(EXIT_FAILURE, "Cannot create mbuf pool\n"); 332 333 tsc_dynfield_offset = 334 rte_mbuf_dynfield_register(&tsc_dynfield_desc); 335 if (tsc_dynfield_offset < 0) 336 rte_exit(EXIT_FAILURE, "Cannot register mbuf field\n"); 337 338 /* initialize all ports */ 339 RTE_ETH_FOREACH_DEV(portid) 340 if (port_init(portid, mbuf_pool) != 0) 341 rte_exit(EXIT_FAILURE, "Cannot init port %"PRIu16"\n", 342 portid); 343 344 if (rte_lcore_count() > 1) 345 printf("\nWARNING: Too much enabled lcores - " 346 "App uses only 1 lcore\n"); 347 348 /* call lcore_main on main core only */ 349 lcore_main(); 350 351 /* clean up the EAL */ 352 rte_eal_cleanup(); 353 354 return 0; 355 } 356