samples/bpf/xdpsock_user.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright(c) 2017 - 2018 Intel Corporation. */
   3
   4 #include <asm/barrier.h>
   5 #include <errno.h>
   6 #include <getopt.h>
   7 #include <libgen.h>
   8 #include <linux/bpf.h>
   9 #include <linux/compiler.h>
  10 #include <linux/if_link.h>
  11 #include <linux/if_xdp.h>
  12 #include <linux/if_ether.h>
  13 #include <locale.h>
  14 #include <net/ethernet.h>
  15 #include <net/if.h>
  16 #include <poll.h>
  17 #include <pthread.h>
  18 #include <signal.h>
  19 #include <stdbool.h>
  20 #include <stdio.h>
  21 #include <stdlib.h>
  22 #include <string.h>
  23 #include <sys/mman.h>
  24 #include <sys/resource.h>
  25 #include <sys/socket.h>
  26 #include <sys/types.h>
  27 #include <time.h>
  28 #include <unistd.h>
  29
  30 #include "libbpf.h"
  31 #include "xsk.h"
  32 #include <bpf/bpf.h>
  33
  34 #ifndef SOL_XDP
  35 #define SOL_XDP 283
  36 #endif
  37
  38 #ifndef AF_XDP
  39 #define AF_XDP 44
  40 #endif
  41
  42 #ifndef PF_XDP
  43 #define PF_XDP AF_XDP
  44 #endif
  45
  46 #define NUM_FRAMES (4 * 1024)
  47 #define BATCH_SIZE 64
  48
  49 #define DEBUG_HEXDUMP 0
  50 #define MAX_SOCKS 8
  51
  52 typedef __u64 u64;
  53 typedef __u32 u32;
  54
  55 static unsigned long prev_time;
  56
  57 enum benchmark_type {
  58         BENCH_RXDROP = 0,
  59         BENCH_TXONLY = 1,
  60         BENCH_L2FWD = 2,
  61 };
  62
  63 static enum benchmark_type opt_bench = BENCH_RXDROP;
  64 static u32 opt_xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
  65 static const char *opt_if = "";
  66 static int opt_ifindex;
  67 static int opt_queue;
  68 static int opt_poll;
  69 static int opt_interval = 1;
  70 static u32 opt_xdp_bind_flags = XDP_USE_NEED_WAKEUP;
  71 static u32 opt_umem_flags;
  72 static int opt_unaligned_chunks;
  73 static int opt_mmap_flags;
  74 static u32 opt_xdp_bind_flags;
  75 static int opt_xsk_frame_size = XSK_UMEM__DEFAULT_FRAME_SIZE;
  76 static int opt_timeout = 1000;
  77 static bool opt_need_wakeup = true;
  78 static __u32 prog_id;
  79
  80 struct xsk_umem_info {
  81         struct xsk_ring_prod fq;
  82         struct xsk_ring_cons cq;
  83         struct xsk_umem *umem;
  84         void *buffer;
  85 };
  86
  87 struct xsk_socket_info {
  88         struct xsk_ring_cons rx;
  89         struct xsk_ring_prod tx;
  90         struct xsk_umem_info *umem;
  91         struct xsk_socket *xsk;
  92         unsigned long rx_npkts;
  93         unsigned long tx_npkts;
  94         unsigned long prev_rx_npkts;
  95         unsigned long prev_tx_npkts;
  96         u32 outstanding_tx;
  97 };
  98
  99 static int num_socks;
 100 struct xsk_socket_info *xsks[MAX_SOCKS];
 101
 102 static unsigned long get_nsecs(void)
 103 {
 104         struct timespec ts;
 105
 106         clock_gettime(CLOCK_MONOTONIC, &ts);
 107         return ts.tv_sec * 1000000000UL + ts.tv_nsec;
 108 }
 109
 110 static void print_benchmark(bool running)
 111 {
 112         const char *bench_str = "INVALID";
 113
 114         if (opt_bench == BENCH_RXDROP)
 115                 bench_str = "rxdrop";
 116         else if (opt_bench == BENCH_TXONLY)
 117                 bench_str = "txonly";
 118         else if (opt_bench == BENCH_L2FWD)
 119                 bench_str = "l2fwd";
 120
 121         printf("%s:%d %s ", opt_if, opt_queue, bench_str);
 122         if (opt_xdp_flags & XDP_FLAGS_SKB_MODE)
 123                 printf("xdp-skb ");
 124         else if (opt_xdp_flags & XDP_FLAGS_DRV_MODE)
 125                 printf("xdp-drv ");
 126         else
 127                 printf("        ");
 128
 129         if (opt_poll)
 130                 printf("poll() ");
 131
 132         if (running) {
 133                 printf("running...");
 134                 fflush(stdout);
 135         }
 136 }
 137
 138 static void dump_stats(void)
 139 {
 140         unsigned long now = get_nsecs();
 141         long dt = now - prev_time;
 142         int i;
 143
 144         prev_time = now;
 145
 146         for (i = 0; i < num_socks && xsks[i]; i++) {
 147                 char *fmt = "%-15s %'-11.0f %'-11lu\n";
 148                 double rx_pps, tx_pps;
 149
 150                 rx_pps = (xsks[i]->rx_npkts - xsks[i]->prev_rx_npkts) *
 151                          1000000000. / dt;
 152                 tx_pps = (xsks[i]->tx_npkts - xsks[i]->prev_tx_npkts) *
 153                          1000000000. / dt;
 154
 155                 printf("\n sock%d@", i);
 156                 print_benchmark(false);
 157                 printf("\n");
 158
 159                 printf("%-15s %-11s %-11s %-11.2f\n", "", "pps", "pkts",
 160                        dt / 1000000000.);
 161                 printf(fmt, "rx", rx_pps, xsks[i]->rx_npkts);
 162                 printf(fmt, "tx", tx_pps, xsks[i]->tx_npkts);
 163
 164                 xsks[i]->prev_rx_npkts = xsks[i]->rx_npkts;
 165                 xsks[i]->prev_tx_npkts = xsks[i]->tx_npkts;
 166         }
 167 }
 168
 169 static void *poller(void *arg)
 170 {
 171         (void)arg;
 172         for (;;) {
 173                 sleep(opt_interval);
 174                 dump_stats();
 175         }
 176
 177         return NULL;
 178 }
 179
 180 static void remove_xdp_program(void)
 181 {
 182         __u32 curr_prog_id = 0;
 183
 184         if (bpf_get_link_xdp_id(opt_ifindex, &curr_prog_id, opt_xdp_flags)) {
 185                 printf("bpf_get_link_xdp_id failed\n");
 186                 exit(EXIT_FAILURE);
 187         }
 188         if (prog_id == curr_prog_id)
 189                 bpf_set_link_xdp_fd(opt_ifindex, -1, opt_xdp_flags);
 190         else if (!curr_prog_id)
 191                 printf("couldn't find a prog id on a given interface\n");
 192         else
 193                 printf("program on interface changed, not removing\n");
 194 }
 195
 196 static void int_exit(int sig)
 197 {
 198         struct xsk_umem *umem = xsks[0]->umem->umem;
 199
 200         (void)sig;
 201
 202         dump_stats();
 203         xsk_socket__delete(xsks[0]->xsk);
 204         (void)xsk_umem__delete(umem);
 205         remove_xdp_program();
 206
 207         exit(EXIT_SUCCESS);
 208 }
 209
 210 static void __exit_with_error(int error, const char *file, const char *func,
 211                               int line)
 212 {
 213         fprintf(stderr, "%s:%s:%i: errno: %d/\"%s\"\n", file, func,
 214                 line, error, strerror(error));
 215         dump_stats();
 216         remove_xdp_program();
 217         exit(EXIT_FAILURE);
 218 }
 219
 220 #define exit_with_error(error) __exit_with_error(error, __FILE__, __func__, \
 221                                                  __LINE__)
 222
 223 static const char pkt_data[] =
 224         "\x3c\xfd\xfe\x9e\x7f\x71\xec\xb1\xd7\x98\x3a\xc0\x08\x00\x45\x00"
 225         "\x00\x2e\x00\x00\x00\x00\x40\x11\x88\x97\x05\x08\x07\x08\xc8\x14"
 226         "\x1e\x04\x10\x92\x10\x92\x00\x1a\x6d\xa3\x34\x33\x1f\x69\x40\x6b"
 227         "\x54\x59\xb6\x14\x2d\x11\x44\xbf\xaf\xd9\xbe\xaa";
 228
 229 static void swap_mac_addresses(void *data)
 230 {
 231         struct ether_header *eth = (struct ether_header *)data;
 232         struct ether_addr *src_addr = (struct ether_addr *)&eth->ether_shost;
 233         struct ether_addr *dst_addr = (struct ether_addr *)&eth->ether_dhost;
 234         struct ether_addr tmp;
 235
 236         tmp = *src_addr;
 237         *src_addr = *dst_addr;
 238         *dst_addr = tmp;
 239 }
 240
 241 static void hex_dump(void *pkt, size_t length, u64 addr)
 242 {
 243         const unsigned char *address = (unsigned char *)pkt;
 244         const unsigned char *line = address;
 245         size_t line_size = 32;
 246         unsigned char c;
 247         char buf[32];
 248         int i = 0;
 249
 250         if (!DEBUG_HEXDUMP)
 251                 return;
 252
 253         sprintf(buf, "addr=%llu", addr);
 254         printf("length = %zu\n", length);
 255         printf("%s | ", buf);
 256         while (length-- > 0) {
 257                 printf("%02X ", *address++);
 258                 if (!(++i % line_size) || (length == 0 && i % line_size)) {
 259                         if (length == 0) {
 260                                 while (i++ % line_size)
 261                                         printf("__ ");
 262                         }
 263                         printf(" | ");  /* right close */
 264                         while (line < address) {
 265                                 c = *line++;
 266                                 printf("%c", (c < 33 || c == 255) ? 0x2E : c);
 267                         }
 268                         printf("\n");
 269                         if (length > 0)
 270                                 printf("%s | ", buf);
 271                 }
 272         }
 273         printf("\n");
 274 }
 275
 276 static size_t gen_eth_frame(struct xsk_umem_info *umem, u64 addr)
 277 {
 278         memcpy(xsk_umem__get_data(umem->buffer, addr), pkt_data,
 279                sizeof(pkt_data) - 1);
 280         return sizeof(pkt_data) - 1;
 281 }
 282
 283 static struct xsk_umem_info *xsk_configure_umem(void *buffer, u64 size)
 284 {
 285         struct xsk_umem_info *umem;
 286         struct xsk_umem_config cfg = {
 287                 .fill_size = XSK_RING_PROD__DEFAULT_NUM_DESCS,
 288                 .comp_size = XSK_RING_CONS__DEFAULT_NUM_DESCS,
 289                 .frame_size = opt_xsk_frame_size,
 290                 .frame_headroom = XSK_UMEM__DEFAULT_FRAME_HEADROOM,
 291                 .flags = opt_umem_flags
 292         };
 293
 294         int ret;
 295
 296         umem = calloc(1, sizeof(*umem));
 297         if (!umem)
 298                 exit_with_error(errno);
 299
 300         ret = xsk_umem__create(&umem->umem, buffer, size, &umem->fq, &umem->cq,
 301                                &cfg);
 302
 303         if (ret)
 304                 exit_with_error(-ret);
 305
 306         umem->buffer = buffer;
 307         return umem;
 308 }
 309
 310 static struct xsk_socket_info *xsk_configure_socket(struct xsk_umem_info *umem)
 311 {
 312         struct xsk_socket_config cfg;
 313         struct xsk_socket_info *xsk;
 314         int ret;
 315         u32 idx;
 316         int i;
 317
 318         xsk = calloc(1, sizeof(*xsk));
 319         if (!xsk)
 320                 exit_with_error(errno);
 321
 322         xsk->umem = umem;
 323         cfg.rx_size = XSK_RING_CONS__DEFAULT_NUM_DESCS;
 324         cfg.tx_size = XSK_RING_PROD__DEFAULT_NUM_DESCS;
 325         cfg.libbpf_flags = 0;
 326         cfg.xdp_flags = opt_xdp_flags;
 327         cfg.bind_flags = opt_xdp_bind_flags;
 328         ret = xsk_socket__create(&xsk->xsk, opt_if, opt_queue, umem->umem,
 329                                  &xsk->rx, &xsk->tx, &cfg);
 330         if (ret)
 331                 exit_with_error(-ret);
 332
 333         ret = bpf_get_link_xdp_id(opt_ifindex, &prog_id, opt_xdp_flags);
 334         if (ret)
 335                 exit_with_error(-ret);
 336
 337         ret = xsk_ring_prod__reserve(&xsk->umem->fq,
 338                                      XSK_RING_PROD__DEFAULT_NUM_DESCS,
 339                                      &idx);
 340         if (ret != XSK_RING_PROD__DEFAULT_NUM_DESCS)
 341                 exit_with_error(-ret);
 342         for (i = 0; i < XSK_RING_PROD__DEFAULT_NUM_DESCS; i++)
 343                 *xsk_ring_prod__fill_addr(&xsk->umem->fq, idx++) =
 344                         i * opt_xsk_frame_size;
 345         xsk_ring_prod__submit(&xsk->umem->fq,
 346                               XSK_RING_PROD__DEFAULT_NUM_DESCS);
 347
 348         return xsk;
 349 }
 350
 351 static struct option long_options[] = {
 352         {"rxdrop", no_argument, 0, 'r'},
 353         {"txonly", no_argument, 0, 't'},
 354         {"l2fwd", no_argument, 0, 'l'},
 355         {"interface", required_argument, 0, 'i'},
 356         {"queue", required_argument, 0, 'q'},
 357         {"poll", no_argument, 0, 'p'},
 358         {"xdp-skb", no_argument, 0, 'S'},
 359         {"xdp-native", no_argument, 0, 'N'},
 360         {"interval", required_argument, 0, 'n'},
 361         {"zero-copy", no_argument, 0, 'z'},
 362         {"copy", no_argument, 0, 'c'},
 363         {"frame-size", required_argument, 0, 'f'},
 364         {"no-need-wakeup", no_argument, 0, 'm'},
 365         {"unaligned", no_argument, 0, 'u'},
 366         {0, 0, 0, 0}
 367 };
 368
 369 static void usage(const char *prog)
 370 {
 371         const char *str =
 372                 "  Usage: %s [OPTIONS]\n"
 373                 "  Options:\n"
 374                 "  -r, --rxdrop         Discard all incoming packets (default)\n"
 375                 "  -t, --txonly         Only send packets\n"
 376                 "  -l, --l2fwd          MAC swap L2 forwarding\n"
 377                 "  -i, --interface=n    Run on interface n\n"
 378                 "  -q, --queue=n        Use queue n (default 0)\n"
 379                 "  -p, --poll           Use poll syscall\n"
 380                 "  -S, --xdp-skb=n      Use XDP skb-mod\n"
 381                 "  -N, --xdp-native=n   Enfore XDP native mode\n"
 382                 "  -n, --interval=n     Specify statistics update interval (default 1 sec).\n"
 383                 "  -z, --zero-copy      Force zero-copy mode.\n"
 384                 "  -c, --copy           Force copy mode.\n"
 385                 "  -f, --frame-size=n   Set the frame size (must be a power of two, default is %d).\n"
 386                 "  -m, --no-need-wakeup Turn off use of driver need wakeup flag.\n"
 387                 "  -f, --frame-size=n   Set the frame size (must be a power of two in aligned mode, default is %d).\n"
 388                 "  -u, --unaligned      Enable unaligned chunk placement\n"
 389                 "\n";
 390         fprintf(stderr, str, prog, XSK_UMEM__DEFAULT_FRAME_SIZE);
 391         exit(EXIT_FAILURE);
 392 }
 393
 394 static void parse_command_line(int argc, char **argv)
 395 {
 396         int option_index, c;
 397
 398         opterr = 0;
 399
 400         for (;;) {
 401                 c = getopt_long(argc, argv, "Frtli:q:psSNn:czf:mu",
 402                                 long_options, &option_index);
 403                 if (c == -1)
 404                         break;
 405
 406                 switch (c) {
 407                 case 'r':
 408                         opt_bench = BENCH_RXDROP;
 409                         break;
 410                 case 't':
 411                         opt_bench = BENCH_TXONLY;
 412                         break;
 413                 case 'l':
 414                         opt_bench = BENCH_L2FWD;
 415                         break;
 416                 case 'i':
 417                         opt_if = optarg;
 418                         break;
 419                 case 'q':
 420                         opt_queue = atoi(optarg);
 421                         break;
 422                 case 'p':
 423                         opt_poll = 1;
 424                         break;
 425                 case 'S':
 426                         opt_xdp_flags |= XDP_FLAGS_SKB_MODE;
 427                         opt_xdp_bind_flags |= XDP_COPY;
 428                         break;
 429                 case 'N':
 430                         opt_xdp_flags |= XDP_FLAGS_DRV_MODE;
 431                         break;
 432                 case 'n':
 433                         opt_interval = atoi(optarg);
 434                         break;
 435                 case 'z':
 436                         opt_xdp_bind_flags |= XDP_ZEROCOPY;
 437                         break;
 438                 case 'c':
 439                         opt_xdp_bind_flags |= XDP_COPY;
 440                         break;
 441                 case 'u':
 442                         opt_umem_flags |= XDP_UMEM_UNALIGNED_CHUNK_FLAG;
 443                         opt_unaligned_chunks = 1;
 444                         opt_mmap_flags = MAP_HUGETLB;
 445                         break;
 446                 case 'F':
 447                         opt_xdp_flags &= ~XDP_FLAGS_UPDATE_IF_NOEXIST;
 448                         break;
 449                 case 'f':
 450                         opt_xsk_frame_size = atoi(optarg);
 451                 case 'm':
 452                         opt_need_wakeup = false;
 453                         opt_xdp_bind_flags &= ~XDP_USE_NEED_WAKEUP;
 454                         break;
 455
 456                 default:
 457                         usage(basename(argv[0]));
 458                 }
 459         }
 460
 461         opt_ifindex = if_nametoindex(opt_if);
 462         if (!opt_ifindex) {
 463                 fprintf(stderr, "ERROR: interface \"%s\" does not exist\n",
 464                         opt_if);
 465                 usage(basename(argv[0]));
 466         }
 467
 468         if ((opt_xsk_frame_size & (opt_xsk_frame_size - 1)) &&
 469             !opt_unaligned_chunks) {
 470                 fprintf(stderr, "--frame-size=%d is not a power of two\n",
 471                         opt_xsk_frame_size);
 472                 usage(basename(argv[0]));
 473         }
 474 }
 475
 476 static void kick_tx(struct xsk_socket_info *xsk)
 477 {
 478         int ret;
 479
 480         ret = sendto(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, 0);
 481         if (ret >= 0 || errno == ENOBUFS || errno == EAGAIN || errno == EBUSY)
 482                 return;
 483         exit_with_error(errno);
 484 }
 485
 486 static inline void complete_tx_l2fwd(struct xsk_socket_info *xsk,
 487                                      struct pollfd *fds)
 488 {
 489         struct xsk_umem_info *umem = xsk->umem;
 490         u32 idx_cq = 0, idx_fq = 0;
 491         unsigned int rcvd;
 492         size_t ndescs;
 493
 494         if (!xsk->outstanding_tx)
 495                 return;
 496
 497         if (!opt_need_wakeup || xsk_ring_prod__needs_wakeup(&xsk->tx))
 498                 kick_tx(xsk);
 499
 500         ndescs = (xsk->outstanding_tx > BATCH_SIZE) ? BATCH_SIZE :
 501                 xsk->outstanding_tx;
 502
 503         /* re-add completed Tx buffers */
 504         rcvd = xsk_ring_cons__peek(&umem->cq, ndescs, &idx_cq);
 505         if (rcvd > 0) {
 506                 unsigned int i;
 507                 int ret;
 508
 509                 ret = xsk_ring_prod__reserve(&umem->fq, rcvd, &idx_fq);
 510                 while (ret != rcvd) {
 511                         if (ret < 0)
 512                                 exit_with_error(-ret);
 513                         if (xsk_ring_prod__needs_wakeup(&umem->fq))
 514                                 ret = poll(fds, num_socks, opt_timeout);
 515                         ret = xsk_ring_prod__reserve(&umem->fq, rcvd, &idx_fq);
 516                 }
 517
 518                 for (i = 0; i < rcvd; i++)
 519                         *xsk_ring_prod__fill_addr(&umem->fq, idx_fq++) =
 520                                 *xsk_ring_cons__comp_addr(&umem->cq, idx_cq++);
 521
 522                 xsk_ring_prod__submit(&xsk->umem->fq, rcvd);
 523                 xsk_ring_cons__release(&xsk->umem->cq, rcvd);
 524                 xsk->outstanding_tx -= rcvd;
 525                 xsk->tx_npkts += rcvd;
 526         }
 527 }
 528
 529 static inline void complete_tx_only(struct xsk_socket_info *xsk)
 530 {
 531         unsigned int rcvd;
 532         u32 idx;
 533
 534         if (!xsk->outstanding_tx)
 535                 return;
 536
 537         if (!opt_need_wakeup || xsk_ring_prod__needs_wakeup(&xsk->tx))
 538                 kick_tx(xsk);
 539
 540         rcvd = xsk_ring_cons__peek(&xsk->umem->cq, BATCH_SIZE, &idx);
 541         if (rcvd > 0) {
 542                 xsk_ring_cons__release(&xsk->umem->cq, rcvd);
 543                 xsk->outstanding_tx -= rcvd;
 544                 xsk->tx_npkts += rcvd;
 545         }
 546 }
 547
 548 static void rx_drop(struct xsk_socket_info *xsk, struct pollfd *fds)
 549 {
 550         unsigned int rcvd, i;
 551         u32 idx_rx = 0, idx_fq = 0;
 552         int ret;
 553
 554         rcvd = xsk_ring_cons__peek(&xsk->rx, BATCH_SIZE, &idx_rx);
 555         if (!rcvd) {
 556                 if (xsk_ring_prod__needs_wakeup(&xsk->umem->fq))
 557                         ret = poll(fds, num_socks, opt_timeout);
 558                 return;
 559         }
 560
 561         ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
 562         while (ret != rcvd) {
 563                 if (ret < 0)
 564                         exit_with_error(-ret);
 565                 if (xsk_ring_prod__needs_wakeup(&xsk->umem->fq))
 566                         ret = poll(fds, num_socks, opt_timeout);
 567                 ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
 568         }
 569
 570         for (i = 0; i < rcvd; i++) {
 571                 u64 addr = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx)->addr;
 572                 u32 len = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx++)->len;
 573                 u64 orig = xsk_umem__extract_addr(addr);
 574
 575                 addr = xsk_umem__add_offset_to_addr(addr);
 576                 char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
 577
 578                 hex_dump(pkt, len, addr);
 579                 *xsk_ring_prod__fill_addr(&xsk->umem->fq, idx_fq++) = orig;
 580         }
 581
 582         xsk_ring_prod__submit(&xsk->umem->fq, rcvd);
 583         xsk_ring_cons__release(&xsk->rx, rcvd);
 584         xsk->rx_npkts += rcvd;
 585 }
 586
 587 static void rx_drop_all(void)
 588 {
 589         struct pollfd fds[MAX_SOCKS + 1];
 590         int i, ret;
 591
 592         memset(fds, 0, sizeof(fds));
 593
 594         for (i = 0; i < num_socks; i++) {
 595                 fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
 596                 fds[i].events = POLLIN;
 597         }
 598
 599         for (;;) {
 600                 if (opt_poll) {
 601                         ret = poll(fds, num_socks, opt_timeout);
 602                         if (ret <= 0)
 603                                 continue;
 604                 }
 605
 606                 for (i = 0; i < num_socks; i++)
 607                         rx_drop(xsks[i], fds);
 608         }
 609 }
 610
 611 static void tx_only(struct xsk_socket_info *xsk, u32 frame_nb)
 612 {
 613         u32 idx;
 614
 615         if (xsk_ring_prod__reserve(&xsk->tx, BATCH_SIZE, &idx) == BATCH_SIZE) {
 616                 unsigned int i;
 617
 618                 for (i = 0; i < BATCH_SIZE; i++) {
 619                         xsk_ring_prod__tx_desc(&xsk->tx, idx + i)->addr =
 620                                 (frame_nb + i) << XSK_UMEM__DEFAULT_FRAME_SHIFT;
 621                         xsk_ring_prod__tx_desc(&xsk->tx, idx + i)->len =
 622                                 sizeof(pkt_data) - 1;
 623                 }
 624
 625                 xsk_ring_prod__submit(&xsk->tx, BATCH_SIZE);
 626                 xsk->outstanding_tx += BATCH_SIZE;
 627                 frame_nb += BATCH_SIZE;
 628                 frame_nb %= NUM_FRAMES;
 629         }
 630
 631         complete_tx_only(xsk);
 632 }
 633
 634 static void tx_only_all(void)
 635 {
 636         struct pollfd fds[MAX_SOCKS];
 637         u32 frame_nb[MAX_SOCKS] = {};
 638         int i, ret;
 639
 640         memset(fds, 0, sizeof(fds));
 641         for (i = 0; i < num_socks; i++) {
 642                 fds[0].fd = xsk_socket__fd(xsks[i]->xsk);
 643                 fds[0].events = POLLOUT;
 644         }
 645
 646         for (;;) {
 647                 if (opt_poll) {
 648                         ret = poll(fds, num_socks, opt_timeout);
 649                         if (ret <= 0)
 650                                 continue;
 651
 652                         if (!(fds[0].revents & POLLOUT))
 653                                 continue;
 654                 }
 655
 656                 for (i = 0; i < num_socks; i++)
 657                         tx_only(xsks[i], frame_nb[i]);
 658         }
 659 }
 660
 661 static void l2fwd(struct xsk_socket_info *xsk, struct pollfd *fds)
 662 {
 663         unsigned int rcvd, i;
 664         u32 idx_rx = 0, idx_tx = 0;
 665         int ret;
 666
 667         complete_tx_l2fwd(xsk, fds);
 668
 669         rcvd = xsk_ring_cons__peek(&xsk->rx, BATCH_SIZE, &idx_rx);
 670         if (!rcvd) {
 671                 if (xsk_ring_prod__needs_wakeup(&xsk->umem->fq))
 672                         ret = poll(fds, num_socks, opt_timeout);
 673                 return;
 674         }
 675
 676         ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
 677         while (ret != rcvd) {
 678                 if (ret < 0)
 679                         exit_with_error(-ret);
 680                 if (xsk_ring_prod__needs_wakeup(&xsk->tx))
 681                         kick_tx(xsk);
 682                 ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
 683         }
 684
 685         for (i = 0; i < rcvd; i++) {
 686                 u64 addr = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx)->addr;
 687                 u32 len = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx++)->len;
 688                 u64 orig = addr;
 689
 690                 addr = xsk_umem__add_offset_to_addr(addr);
 691                 char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
 692
 693                 swap_mac_addresses(pkt);
 694
 695                 hex_dump(pkt, len, addr);
 696                 xsk_ring_prod__tx_desc(&xsk->tx, idx_tx)->addr = orig;
 697                 xsk_ring_prod__tx_desc(&xsk->tx, idx_tx++)->len = len;
 698         }
 699
 700         xsk_ring_prod__submit(&xsk->tx, rcvd);
 701         xsk_ring_cons__release(&xsk->rx, rcvd);
 702
 703         xsk->rx_npkts += rcvd;
 704         xsk->outstanding_tx += rcvd;
 705 }
 706
 707 static void l2fwd_all(void)
 708 {
 709         struct pollfd fds[MAX_SOCKS];
 710         int i, ret;
 711
 712         memset(fds, 0, sizeof(fds));
 713
 714         for (i = 0; i < num_socks; i++) {
 715                 fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
 716                 fds[i].events = POLLOUT | POLLIN;
 717         }
 718
 719         for (;;) {
 720                 if (opt_poll) {
 721                         ret = poll(fds, num_socks, opt_timeout);
 722                         if (ret <= 0)
 723                                 continue;
 724                 }
 725
 726                 for (i = 0; i < num_socks; i++)
 727                         l2fwd(xsks[i], fds);
 728         }
 729 }
 730
 731 int main(int argc, char **argv)
 732 {
 733         struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
 734         struct xsk_umem_info *umem;
 735         pthread_t pt;
 736         void *bufs;
 737         int ret;
 738
 739         parse_command_line(argc, argv);
 740
 741         if (setrlimit(RLIMIT_MEMLOCK, &r)) {
 742                 fprintf(stderr, "ERROR: setrlimit(RLIMIT_MEMLOCK) \"%s\"\n",
 743                         strerror(errno));
 744                 exit(EXIT_FAILURE);
 745         }
 746
 747         /* Reserve memory for the umem. Use hugepages if unaligned chunk mode */
 748         bufs = mmap(NULL, NUM_FRAMES * opt_xsk_frame_size,
 749                     PROT_READ | PROT_WRITE,
 750                     MAP_PRIVATE | MAP_ANONYMOUS | opt_mmap_flags, -1, 0);
 751         if (bufs == MAP_FAILED) {
 752                 printf("ERROR: mmap failed\n");
 753                 exit(EXIT_FAILURE);
 754         }
 755        /* Create sockets... */
 756         umem = xsk_configure_umem(bufs, NUM_FRAMES * opt_xsk_frame_size);
 757         xsks[num_socks++] = xsk_configure_socket(umem);
 758
 759         if (opt_bench == BENCH_TXONLY) {
 760                 int i;
 761
 762                 for (i = 0; i < NUM_FRAMES; i++)
 763                         (void)gen_eth_frame(umem, i * opt_xsk_frame_size);
 764         }
 765
 766         signal(SIGINT, int_exit);
 767         signal(SIGTERM, int_exit);
 768         signal(SIGABRT, int_exit);
 769
 770         setlocale(LC_ALL, "");
 771
 772         ret = pthread_create(&pt, NULL, poller, NULL);
 773         if (ret)
 774                 exit_with_error(ret);
 775
 776         prev_time = get_nsecs();
 777
 778         if (opt_bench == BENCH_RXDROP)
 779                 rx_drop_all();
 780         else if (opt_bench == BENCH_TXONLY)
 781                 tx_only_all();
 782         else
 783                 l2fwd_all();
 784
 785         return 0;
 786 }