samples/bpf/xdpsock_user.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright(c) 2017 - 2018 Intel Corporation. */
   3
   4 #include <assert.h>
   5 #include <errno.h>
   6 #include <getopt.h>
   7 #include <libgen.h>
   8 #include <linux/bpf.h>
   9 #include <linux/if_link.h>
  10 #include <linux/if_xdp.h>
  11 #include <linux/if_ether.h>
  12 #include <net/if.h>
  13 #include <signal.h>
  14 #include <stdbool.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <string.h>
  18 #include <net/ethernet.h>
  19 #include <sys/resource.h>
  20 #include <sys/socket.h>
  21 #include <sys/mman.h>
  22 #include <time.h>
  23 #include <unistd.h>
  24 #include <pthread.h>
  25 #include <locale.h>
  26 #include <sys/types.h>
  27 #include <poll.h>
  28
  29 #include "bpf_load.h"
  30 #include "bpf_util.h"
  31 #include <bpf/bpf.h>
  32
  33 #include "xdpsock.h"
  34
  35 #ifndef SOL_XDP
  36 #define SOL_XDP 283
  37 #endif
  38
  39 #ifndef AF_XDP
  40 #define AF_XDP 44
  41 #endif
  42
  43 #ifndef PF_XDP
  44 #define PF_XDP AF_XDP
  45 #endif
  46
  47 #define NUM_FRAMES 131072
  48 #define FRAME_HEADROOM 0
  49 #define FRAME_SHIFT 11
  50 #define FRAME_SIZE 2048
  51 #define NUM_DESCS 1024
  52 #define BATCH_SIZE 16
  53
  54 #define FQ_NUM_DESCS 1024
  55 #define CQ_NUM_DESCS 1024
  56
  57 #define DEBUG_HEXDUMP 0
  58
  59 typedef __u64 u64;
  60 typedef __u32 u32;
  61
  62 static unsigned long prev_time;
  63
  64 enum benchmark_type {
  65         BENCH_RXDROP = 0,
  66         BENCH_TXONLY = 1,
  67         BENCH_L2FWD = 2,
  68 };
  69
  70 static enum benchmark_type opt_bench = BENCH_RXDROP;
  71 static u32 opt_xdp_flags;
  72 static const char *opt_if = "";
  73 static int opt_ifindex;
  74 static int opt_queue;
  75 static int opt_poll;
  76 static int opt_shared_packet_buffer;
  77 static int opt_interval = 1;
  78 static u32 opt_xdp_bind_flags;
  79
  80 struct xdp_umem_uqueue {
  81         u32 cached_prod;
  82         u32 cached_cons;
  83         u32 mask;
  84         u32 size;
  85         u32 *producer;
  86         u32 *consumer;
  87         u64 *ring;
  88         void *map;
  89 };
  90
  91 struct xdp_umem {
  92         char *frames;
  93         struct xdp_umem_uqueue fq;
  94         struct xdp_umem_uqueue cq;
  95         int fd;
  96 };
  97
  98 struct xdp_uqueue {
  99         u32 cached_prod;
 100         u32 cached_cons;
 101         u32 mask;
 102         u32 size;
 103         u32 *producer;
 104         u32 *consumer;
 105         struct xdp_desc *ring;
 106         void *map;
 107 };
 108
 109 struct xdpsock {
 110         struct xdp_uqueue rx;
 111         struct xdp_uqueue tx;
 112         int sfd;
 113         struct xdp_umem *umem;
 114         u32 outstanding_tx;
 115         unsigned long rx_npkts;
 116         unsigned long tx_npkts;
 117         unsigned long prev_rx_npkts;
 118         unsigned long prev_tx_npkts;
 119 };
 120
 121 #define MAX_SOCKS 4
 122 static int num_socks;
 123 struct xdpsock *xsks[MAX_SOCKS];
 124
 125 static unsigned long get_nsecs(void)
 126 {
 127         struct timespec ts;
 128
 129         clock_gettime(CLOCK_MONOTONIC, &ts);
 130         return ts.tv_sec * 1000000000UL + ts.tv_nsec;
 131 }
 132
 133 static void dump_stats(void);
 134
 135 #define lassert(expr)                                                   \
 136         do {                                                            \
 137                 if (!(expr)) {                                          \
 138                         fprintf(stderr, "%s:%s:%i: Assertion failed: "  \
 139                                 #expr ": errno: %d/\"%s\"\n",           \
 140                                 __FILE__, __func__, __LINE__,           \
 141                                 errno, strerror(errno));                \
 142                         dump_stats();                                   \
 143                         exit(EXIT_FAILURE);                             \
 144                 }                                                       \
 145         } while (0)
 146
 147 #define barrier() __asm__ __volatile__("": : :"memory")
 148 #define u_smp_rmb() barrier()
 149 #define u_smp_wmb() barrier()
 150 #define likely(x) __builtin_expect(!!(x), 1)
 151 #define unlikely(x) __builtin_expect(!!(x), 0)
 152
 153 static const char pkt_data[] =
 154         "\x3c\xfd\xfe\x9e\x7f\x71\xec\xb1\xd7\x98\x3a\xc0\x08\x00\x45\x00"
 155         "\x00\x2e\x00\x00\x00\x00\x40\x11\x88\x97\x05\x08\x07\x08\xc8\x14"
 156         "\x1e\x04\x10\x92\x10\x92\x00\x1a\x6d\xa3\x34\x33\x1f\x69\x40\x6b"
 157         "\x54\x59\xb6\x14\x2d\x11\x44\xbf\xaf\xd9\xbe\xaa";
 158
 159 static inline u32 umem_nb_free(struct xdp_umem_uqueue *q, u32 nb)
 160 {
 161         u32 free_entries = q->cached_cons - q->cached_prod;
 162
 163         if (free_entries >= nb)
 164                 return free_entries;
 165
 166         /* Refresh the local tail pointer */
 167         q->cached_cons = *q->consumer + q->size;
 168
 169         return q->cached_cons - q->cached_prod;
 170 }
 171
 172 static inline u32 xq_nb_free(struct xdp_uqueue *q, u32 ndescs)
 173 {
 174         u32 free_entries = q->cached_cons - q->cached_prod;
 175
 176         if (free_entries >= ndescs)
 177                 return free_entries;
 178
 179         /* Refresh the local tail pointer */
 180         q->cached_cons = *q->consumer + q->size;
 181         return q->cached_cons - q->cached_prod;
 182 }
 183
 184 static inline u32 umem_nb_avail(struct xdp_umem_uqueue *q, u32 nb)
 185 {
 186         u32 entries = q->cached_prod - q->cached_cons;
 187
 188         if (entries == 0) {
 189                 q->cached_prod = *q->producer;
 190                 entries = q->cached_prod - q->cached_cons;
 191         }
 192
 193         return (entries > nb) ? nb : entries;
 194 }
 195
 196 static inline u32 xq_nb_avail(struct xdp_uqueue *q, u32 ndescs)
 197 {
 198         u32 entries = q->cached_prod - q->cached_cons;
 199
 200         if (entries == 0) {
 201                 q->cached_prod = *q->producer;
 202                 entries = q->cached_prod - q->cached_cons;
 203         }
 204
 205         return (entries > ndescs) ? ndescs : entries;
 206 }
 207
 208 static inline int umem_fill_to_kernel_ex(struct xdp_umem_uqueue *fq,
 209                                          struct xdp_desc *d,
 210                                          size_t nb)
 211 {
 212         u32 i;
 213
 214         if (umem_nb_free(fq, nb) < nb)
 215                 return -ENOSPC;
 216
 217         for (i = 0; i < nb; i++) {
 218                 u32 idx = fq->cached_prod++ & fq->mask;
 219
 220                 fq->ring[idx] = d[i].addr;
 221         }
 222
 223         u_smp_wmb();
 224
 225         *fq->producer = fq->cached_prod;
 226
 227         return 0;
 228 }
 229
 230 static inline int umem_fill_to_kernel(struct xdp_umem_uqueue *fq, u64 *d,
 231                                       size_t nb)
 232 {
 233         u32 i;
 234
 235         if (umem_nb_free(fq, nb) < nb)
 236                 return -ENOSPC;
 237
 238         for (i = 0; i < nb; i++) {
 239                 u32 idx = fq->cached_prod++ & fq->mask;
 240
 241                 fq->ring[idx] = d[i];
 242         }
 243
 244         u_smp_wmb();
 245
 246         *fq->producer = fq->cached_prod;
 247
 248         return 0;
 249 }
 250
 251 static inline size_t umem_complete_from_kernel(struct xdp_umem_uqueue *cq,
 252                                                u64 *d, size_t nb)
 253 {
 254         u32 idx, i, entries = umem_nb_avail(cq, nb);
 255
 256         u_smp_rmb();
 257
 258         for (i = 0; i < entries; i++) {
 259                 idx = cq->cached_cons++ & cq->mask;
 260                 d[i] = cq->ring[idx];
 261         }
 262
 263         if (entries > 0) {
 264                 u_smp_wmb();
 265
 266                 *cq->consumer = cq->cached_cons;
 267         }
 268
 269         return entries;
 270 }
 271
 272 static inline void *xq_get_data(struct xdpsock *xsk, u64 addr)
 273 {
 274         return &xsk->umem->frames[addr];
 275 }
 276
 277 static inline int xq_enq(struct xdp_uqueue *uq,
 278                          const struct xdp_desc *descs,
 279                          unsigned int ndescs)
 280 {
 281         struct xdp_desc *r = uq->ring;
 282         unsigned int i;
 283
 284         if (xq_nb_free(uq, ndescs) < ndescs)
 285                 return -ENOSPC;
 286
 287         for (i = 0; i < ndescs; i++) {
 288                 u32 idx = uq->cached_prod++ & uq->mask;
 289
 290                 r[idx].addr = descs[i].addr;
 291                 r[idx].len = descs[i].len;
 292         }
 293
 294         u_smp_wmb();
 295
 296         *uq->producer = uq->cached_prod;
 297         return 0;
 298 }
 299
 300 static inline int xq_enq_tx_only(struct xdp_uqueue *uq,
 301                                  unsigned int id, unsigned int ndescs)
 302 {
 303         struct xdp_desc *r = uq->ring;
 304         unsigned int i;
 305
 306         if (xq_nb_free(uq, ndescs) < ndescs)
 307                 return -ENOSPC;
 308
 309         for (i = 0; i < ndescs; i++) {
 310                 u32 idx = uq->cached_prod++ & uq->mask;
 311
 312                 r[idx].addr     = (id + i) << FRAME_SHIFT;
 313                 r[idx].len      = sizeof(pkt_data) - 1;
 314         }
 315
 316         u_smp_wmb();
 317
 318         *uq->producer = uq->cached_prod;
 319         return 0;
 320 }
 321
 322 static inline int xq_deq(struct xdp_uqueue *uq,
 323                          struct xdp_desc *descs,
 324                          int ndescs)
 325 {
 326         struct xdp_desc *r = uq->ring;
 327         unsigned int idx;
 328         int i, entries;
 329
 330         entries = xq_nb_avail(uq, ndescs);
 331
 332         u_smp_rmb();
 333
 334         for (i = 0; i < entries; i++) {
 335                 idx = uq->cached_cons++ & uq->mask;
 336                 descs[i] = r[idx];
 337         }
 338
 339         if (entries > 0) {
 340                 u_smp_wmb();
 341
 342                 *uq->consumer = uq->cached_cons;
 343         }
 344
 345         return entries;
 346 }
 347
 348 static void swap_mac_addresses(void *data)
 349 {
 350         struct ether_header *eth = (struct ether_header *)data;
 351         struct ether_addr *src_addr = (struct ether_addr *)&eth->ether_shost;
 352         struct ether_addr *dst_addr = (struct ether_addr *)&eth->ether_dhost;
 353         struct ether_addr tmp;
 354
 355         tmp = *src_addr;
 356         *src_addr = *dst_addr;
 357         *dst_addr = tmp;
 358 }
 359
 360 static void hex_dump(void *pkt, size_t length, u64 addr)
 361 {
 362         const unsigned char *address = (unsigned char *)pkt;
 363         const unsigned char *line = address;
 364         size_t line_size = 32;
 365         unsigned char c;
 366         char buf[32];
 367         int i = 0;
 368
 369         if (!DEBUG_HEXDUMP)
 370                 return;
 371
 372         sprintf(buf, "addr=%llu", addr);
 373         printf("length = %zu\n", length);
 374         printf("%s | ", buf);
 375         while (length-- > 0) {
 376                 printf("%02X ", *address++);
 377                 if (!(++i % line_size) || (length == 0 && i % line_size)) {
 378                         if (length == 0) {
 379                                 while (i++ % line_size)
 380                                         printf("__ ");
 381                         }
 382                         printf(" | ");  /* right close */
 383                         while (line < address) {
 384                                 c = *line++;
 385                                 printf("%c", (c < 33 || c == 255) ? 0x2E : c);
 386                         }
 387                         printf("\n");
 388                         if (length > 0)
 389                                 printf("%s | ", buf);
 390                 }
 391         }
 392         printf("\n");
 393 }
 394
 395 static size_t gen_eth_frame(char *frame)
 396 {
 397         memcpy(frame, pkt_data, sizeof(pkt_data) - 1);
 398         return sizeof(pkt_data) - 1;
 399 }
 400
 401 static struct xdp_umem *xdp_umem_configure(int sfd)
 402 {
 403         int fq_size = FQ_NUM_DESCS, cq_size = CQ_NUM_DESCS;
 404         struct xdp_mmap_offsets off;
 405         struct xdp_umem_reg mr;
 406         struct xdp_umem *umem;
 407         socklen_t optlen;
 408         void *bufs;
 409
 410         umem = calloc(1, sizeof(*umem));
 411         lassert(umem);
 412
 413         lassert(posix_memalign(&bufs, getpagesize(), /* PAGE_SIZE aligned */
 414                                NUM_FRAMES * FRAME_SIZE) == 0);
 415
 416         mr.addr = (__u64)bufs;
 417         mr.len = NUM_FRAMES * FRAME_SIZE;
 418         mr.chunk_size = FRAME_SIZE;
 419         mr.headroom = FRAME_HEADROOM;
 420
 421         lassert(setsockopt(sfd, SOL_XDP, XDP_UMEM_REG, &mr, sizeof(mr)) == 0);
 422         lassert(setsockopt(sfd, SOL_XDP, XDP_UMEM_FILL_RING, &fq_size,
 423                            sizeof(int)) == 0);
 424         lassert(setsockopt(sfd, SOL_XDP, XDP_UMEM_COMPLETION_RING, &cq_size,
 425                            sizeof(int)) == 0);
 426
 427         optlen = sizeof(off);
 428         lassert(getsockopt(sfd, SOL_XDP, XDP_MMAP_OFFSETS, &off,
 429                            &optlen) == 0);
 430
 431         umem->fq.map = mmap(0, off.fr.desc +
 432                             FQ_NUM_DESCS * sizeof(u64),
 433                             PROT_READ | PROT_WRITE,
 434                             MAP_SHARED | MAP_POPULATE, sfd,
 435                             XDP_UMEM_PGOFF_FILL_RING);
 436         lassert(umem->fq.map != MAP_FAILED);
 437
 438         umem->fq.mask = FQ_NUM_DESCS - 1;
 439         umem->fq.size = FQ_NUM_DESCS;
 440         umem->fq.producer = umem->fq.map + off.fr.producer;
 441         umem->fq.consumer = umem->fq.map + off.fr.consumer;
 442         umem->fq.ring = umem->fq.map + off.fr.desc;
 443         umem->fq.cached_cons = FQ_NUM_DESCS;
 444
 445         umem->cq.map = mmap(0, off.cr.desc +
 446                              CQ_NUM_DESCS * sizeof(u64),
 447                              PROT_READ | PROT_WRITE,
 448                              MAP_SHARED | MAP_POPULATE, sfd,
 449                              XDP_UMEM_PGOFF_COMPLETION_RING);
 450         lassert(umem->cq.map != MAP_FAILED);
 451
 452         umem->cq.mask = CQ_NUM_DESCS - 1;
 453         umem->cq.size = CQ_NUM_DESCS;
 454         umem->cq.producer = umem->cq.map + off.cr.producer;
 455         umem->cq.consumer = umem->cq.map + off.cr.consumer;
 456         umem->cq.ring = umem->cq.map + off.cr.desc;
 457
 458         umem->frames = bufs;
 459         umem->fd = sfd;
 460
 461         if (opt_bench == BENCH_TXONLY) {
 462                 int i;
 463
 464                 for (i = 0; i < NUM_FRAMES * FRAME_SIZE; i += FRAME_SIZE)
 465                         (void)gen_eth_frame(&umem->frames[i]);
 466         }
 467
 468         return umem;
 469 }
 470
 471 static struct xdpsock *xsk_configure(struct xdp_umem *umem)
 472 {
 473         struct sockaddr_xdp sxdp = {};
 474         struct xdp_mmap_offsets off;
 475         int sfd, ndescs = NUM_DESCS;
 476         struct xdpsock *xsk;
 477         bool shared = true;
 478         socklen_t optlen;
 479         u64 i;
 480
 481         sfd = socket(PF_XDP, SOCK_RAW, 0);
 482         lassert(sfd >= 0);
 483
 484         xsk = calloc(1, sizeof(*xsk));
 485         lassert(xsk);
 486
 487         xsk->sfd = sfd;
 488         xsk->outstanding_tx = 0;
 489
 490         if (!umem) {
 491                 shared = false;
 492                 xsk->umem = xdp_umem_configure(sfd);
 493         } else {
 494                 xsk->umem = umem;
 495         }
 496
 497         lassert(setsockopt(sfd, SOL_XDP, XDP_RX_RING,
 498                            &ndescs, sizeof(int)) == 0);
 499         lassert(setsockopt(sfd, SOL_XDP, XDP_TX_RING,
 500                            &ndescs, sizeof(int)) == 0);
 501         optlen = sizeof(off);
 502         lassert(getsockopt(sfd, SOL_XDP, XDP_MMAP_OFFSETS, &off,
 503                            &optlen) == 0);
 504
 505         /* Rx */
 506         xsk->rx.map = mmap(NULL,
 507                            off.rx.desc +
 508                            NUM_DESCS * sizeof(struct xdp_desc),
 509                            PROT_READ | PROT_WRITE,
 510                            MAP_SHARED | MAP_POPULATE, sfd,
 511                            XDP_PGOFF_RX_RING);
 512         lassert(xsk->rx.map != MAP_FAILED);
 513
 514         if (!shared) {
 515                 for (i = 0; i < NUM_DESCS * FRAME_SIZE; i += FRAME_SIZE)
 516                         lassert(umem_fill_to_kernel(&xsk->umem->fq, &i, 1)
 517                                 == 0);
 518         }
 519
 520         /* Tx */
 521         xsk->tx.map = mmap(NULL,
 522                            off.tx.desc +
 523                            NUM_DESCS * sizeof(struct xdp_desc),
 524                            PROT_READ | PROT_WRITE,
 525                            MAP_SHARED | MAP_POPULATE, sfd,
 526                            XDP_PGOFF_TX_RING);
 527         lassert(xsk->tx.map != MAP_FAILED);
 528
 529         xsk->rx.mask = NUM_DESCS - 1;
 530         xsk->rx.size = NUM_DESCS;
 531         xsk->rx.producer = xsk->rx.map + off.rx.producer;
 532         xsk->rx.consumer = xsk->rx.map + off.rx.consumer;
 533         xsk->rx.ring = xsk->rx.map + off.rx.desc;
 534
 535         xsk->tx.mask = NUM_DESCS - 1;
 536         xsk->tx.size = NUM_DESCS;
 537         xsk->tx.producer = xsk->tx.map + off.tx.producer;
 538         xsk->tx.consumer = xsk->tx.map + off.tx.consumer;
 539         xsk->tx.ring = xsk->tx.map + off.tx.desc;
 540         xsk->tx.cached_cons = NUM_DESCS;
 541
 542         sxdp.sxdp_family = PF_XDP;
 543         sxdp.sxdp_ifindex = opt_ifindex;
 544         sxdp.sxdp_queue_id = opt_queue;
 545
 546         if (shared) {
 547                 sxdp.sxdp_flags = XDP_SHARED_UMEM;
 548                 sxdp.sxdp_shared_umem_fd = umem->fd;
 549         } else {
 550                 sxdp.sxdp_flags = opt_xdp_bind_flags;
 551         }
 552
 553         lassert(bind(sfd, (struct sockaddr *)&sxdp, sizeof(sxdp)) == 0);
 554
 555         return xsk;
 556 }
 557
 558 static void print_benchmark(bool running)
 559 {
 560         const char *bench_str = "INVALID";
 561
 562         if (opt_bench == BENCH_RXDROP)
 563                 bench_str = "rxdrop";
 564         else if (opt_bench == BENCH_TXONLY)
 565                 bench_str = "txonly";
 566         else if (opt_bench == BENCH_L2FWD)
 567                 bench_str = "l2fwd";
 568
 569         printf("%s:%d %s ", opt_if, opt_queue, bench_str);
 570         if (opt_xdp_flags & XDP_FLAGS_SKB_MODE)
 571                 printf("xdp-skb ");
 572         else if (opt_xdp_flags & XDP_FLAGS_DRV_MODE)
 573                 printf("xdp-drv ");
 574         else
 575                 printf("        ");
 576
 577         if (opt_poll)
 578                 printf("poll() ");
 579
 580         if (running) {
 581                 printf("running...");
 582                 fflush(stdout);
 583         }
 584 }
 585
 586 static void dump_stats(void)
 587 {
 588         unsigned long now = get_nsecs();
 589         long dt = now - prev_time;
 590         int i;
 591
 592         prev_time = now;
 593
 594         for (i = 0; i < num_socks; i++) {
 595                 char *fmt = "%-15s %'-11.0f %'-11lu\n";
 596                 double rx_pps, tx_pps;
 597
 598                 rx_pps = (xsks[i]->rx_npkts - xsks[i]->prev_rx_npkts) *
 599                          1000000000. / dt;
 600                 tx_pps = (xsks[i]->tx_npkts - xsks[i]->prev_tx_npkts) *
 601                          1000000000. / dt;
 602
 603                 printf("\n sock%d@", i);
 604                 print_benchmark(false);
 605                 printf("\n");
 606
 607                 printf("%-15s %-11s %-11s %-11.2f\n", "", "pps", "pkts",
 608                        dt / 1000000000.);
 609                 printf(fmt, "rx", rx_pps, xsks[i]->rx_npkts);
 610                 printf(fmt, "tx", tx_pps, xsks[i]->tx_npkts);
 611
 612                 xsks[i]->prev_rx_npkts = xsks[i]->rx_npkts;
 613                 xsks[i]->prev_tx_npkts = xsks[i]->tx_npkts;
 614         }
 615 }
 616
 617 static void *poller(void *arg)
 618 {
 619         (void)arg;
 620         for (;;) {
 621                 sleep(opt_interval);
 622                 dump_stats();
 623         }
 624
 625         return NULL;
 626 }
 627
 628 static void int_exit(int sig)
 629 {
 630         (void)sig;
 631         dump_stats();
 632         bpf_set_link_xdp_fd(opt_ifindex, -1, opt_xdp_flags);
 633         exit(EXIT_SUCCESS);
 634 }
 635
 636 static struct option long_options[] = {
 637         {"rxdrop", no_argument, 0, 'r'},
 638         {"txonly", no_argument, 0, 't'},
 639         {"l2fwd", no_argument, 0, 'l'},
 640         {"interface", required_argument, 0, 'i'},
 641         {"queue", required_argument, 0, 'q'},
 642         {"poll", no_argument, 0, 'p'},
 643         {"shared-buffer", no_argument, 0, 's'},
 644         {"xdp-skb", no_argument, 0, 'S'},
 645         {"xdp-native", no_argument, 0, 'N'},
 646         {"interval", required_argument, 0, 'n'},
 647         {0, 0, 0, 0}
 648 };
 649
 650 static void usage(const char *prog)
 651 {
 652         const char *str =
 653                 "  Usage: %s [OPTIONS]\n"
 654                 "  Options:\n"
 655                 "  -r, --rxdrop         Discard all incoming packets (default)\n"
 656                 "  -t, --txonly         Only send packets\n"
 657                 "  -l, --l2fwd          MAC swap L2 forwarding\n"
 658                 "  -i, --interface=n    Run on interface n\n"
 659                 "  -q, --queue=n        Use queue n (default 0)\n"
 660                 "  -p, --poll           Use poll syscall\n"
 661                 "  -s, --shared-buffer  Use shared packet buffer\n"
 662                 "  -S, --xdp-skb=n      Use XDP skb-mod\n"
 663                 "  -N, --xdp-native=n   Enfore XDP native mode\n"
 664                 "  -n, --interval=n     Specify statistics update interval (default 1 sec).\n"
 665                 "\n";
 666         fprintf(stderr, str, prog);
 667         exit(EXIT_FAILURE);
 668 }
 669
 670 static void parse_command_line(int argc, char **argv)
 671 {
 672         int option_index, c;
 673
 674         opterr = 0;
 675
 676         for (;;) {
 677                 c = getopt_long(argc, argv, "rtli:q:psSNn:", long_options,
 678                                 &option_index);
 679                 if (c == -1)
 680                         break;
 681
 682                 switch (c) {
 683                 case 'r':
 684                         opt_bench = BENCH_RXDROP;
 685                         break;
 686                 case 't':
 687                         opt_bench = BENCH_TXONLY;
 688                         break;
 689                 case 'l':
 690                         opt_bench = BENCH_L2FWD;
 691                         break;
 692                 case 'i':
 693                         opt_if = optarg;
 694                         break;
 695                 case 'q':
 696                         opt_queue = atoi(optarg);
 697                         break;
 698                 case 's':
 699                         opt_shared_packet_buffer = 1;
 700                         break;
 701                 case 'p':
 702                         opt_poll = 1;
 703                         break;
 704                 case 'S':
 705                         opt_xdp_flags |= XDP_FLAGS_SKB_MODE;
 706                         opt_xdp_bind_flags |= XDP_COPY;
 707                         break;
 708                 case 'N':
 709                         opt_xdp_flags |= XDP_FLAGS_DRV_MODE;
 710                         break;
 711                 case 'n':
 712                         opt_interval = atoi(optarg);
 713                         break;
 714                 default:
 715                         usage(basename(argv[0]));
 716                 }
 717         }
 718
 719         opt_ifindex = if_nametoindex(opt_if);
 720         if (!opt_ifindex) {
 721                 fprintf(stderr, "ERROR: interface \"%s\" does not exist\n",
 722                         opt_if);
 723                 usage(basename(argv[0]));
 724         }
 725 }
 726
 727 static void kick_tx(int fd)
 728 {
 729         int ret;
 730
 731         ret = sendto(fd, NULL, 0, MSG_DONTWAIT, NULL, 0);
 732         if (ret >= 0 || errno == ENOBUFS || errno == EAGAIN)
 733                 return;
 734         lassert(0);
 735 }
 736
 737 static inline void complete_tx_l2fwd(struct xdpsock *xsk)
 738 {
 739         u64 descs[BATCH_SIZE];
 740         unsigned int rcvd;
 741         size_t ndescs;
 742
 743         if (!xsk->outstanding_tx)
 744                 return;
 745
 746         kick_tx(xsk->sfd);
 747         ndescs = (xsk->outstanding_tx > BATCH_SIZE) ? BATCH_SIZE :
 748                  xsk->outstanding_tx;
 749
 750         /* re-add completed Tx buffers */
 751         rcvd = umem_complete_from_kernel(&xsk->umem->cq, descs, ndescs);
 752         if (rcvd > 0) {
 753                 umem_fill_to_kernel(&xsk->umem->fq, descs, rcvd);
 754                 xsk->outstanding_tx -= rcvd;
 755                 xsk->tx_npkts += rcvd;
 756         }
 757 }
 758
 759 static inline void complete_tx_only(struct xdpsock *xsk)
 760 {
 761         u64 descs[BATCH_SIZE];
 762         unsigned int rcvd;
 763
 764         if (!xsk->outstanding_tx)
 765                 return;
 766
 767         kick_tx(xsk->sfd);
 768
 769         rcvd = umem_complete_from_kernel(&xsk->umem->cq, descs, BATCH_SIZE);
 770         if (rcvd > 0) {
 771                 xsk->outstanding_tx -= rcvd;
 772                 xsk->tx_npkts += rcvd;
 773         }
 774 }
 775
 776 static void rx_drop(struct xdpsock *xsk)
 777 {
 778         struct xdp_desc descs[BATCH_SIZE];
 779         unsigned int rcvd, i;
 780
 781         rcvd = xq_deq(&xsk->rx, descs, BATCH_SIZE);
 782         if (!rcvd)
 783                 return;
 784
 785         for (i = 0; i < rcvd; i++) {
 786                 char *pkt = xq_get_data(xsk, descs[i].addr);
 787
 788                 hex_dump(pkt, descs[i].len, descs[i].addr);
 789         }
 790
 791         xsk->rx_npkts += rcvd;
 792
 793         umem_fill_to_kernel_ex(&xsk->umem->fq, descs, rcvd);
 794 }
 795
 796 static void rx_drop_all(void)
 797 {
 798         struct pollfd fds[MAX_SOCKS + 1];
 799         int i, ret, timeout, nfds = 1;
 800
 801         memset(fds, 0, sizeof(fds));
 802
 803         for (i = 0; i < num_socks; i++) {
 804                 fds[i].fd = xsks[i]->sfd;
 805                 fds[i].events = POLLIN;
 806                 timeout = 1000; /* 1sn */
 807         }
 808
 809         for (;;) {
 810                 if (opt_poll) {
 811                         ret = poll(fds, nfds, timeout);
 812                         if (ret <= 0)
 813                                 continue;
 814                 }
 815
 816                 for (i = 0; i < num_socks; i++)
 817                         rx_drop(xsks[i]);
 818         }
 819 }
 820
 821 static void tx_only(struct xdpsock *xsk)
 822 {
 823         int timeout, ret, nfds = 1;
 824         struct pollfd fds[nfds + 1];
 825         unsigned int idx = 0;
 826
 827         memset(fds, 0, sizeof(fds));
 828         fds[0].fd = xsk->sfd;
 829         fds[0].events = POLLOUT;
 830         timeout = 1000; /* 1sn */
 831
 832         for (;;) {
 833                 if (opt_poll) {
 834                         ret = poll(fds, nfds, timeout);
 835                         if (ret <= 0)
 836                                 continue;
 837
 838                         if (fds[0].fd != xsk->sfd ||
 839                             !(fds[0].revents & POLLOUT))
 840                                 continue;
 841                 }
 842
 843                 if (xq_nb_free(&xsk->tx, BATCH_SIZE) >= BATCH_SIZE) {
 844                         lassert(xq_enq_tx_only(&xsk->tx, idx, BATCH_SIZE) == 0);
 845
 846                         xsk->outstanding_tx += BATCH_SIZE;
 847                         idx += BATCH_SIZE;
 848                         idx %= NUM_FRAMES;
 849                 }
 850
 851                 complete_tx_only(xsk);
 852         }
 853 }
 854
 855 static void l2fwd(struct xdpsock *xsk)
 856 {
 857         for (;;) {
 858                 struct xdp_desc descs[BATCH_SIZE];
 859                 unsigned int rcvd, i;
 860                 int ret;
 861
 862                 for (;;) {
 863                         complete_tx_l2fwd(xsk);
 864
 865                         rcvd = xq_deq(&xsk->rx, descs, BATCH_SIZE);
 866                         if (rcvd > 0)
 867                                 break;
 868                 }
 869
 870                 for (i = 0; i < rcvd; i++) {
 871                         char *pkt = xq_get_data(xsk, descs[i].addr);
 872
 873                         swap_mac_addresses(pkt);
 874
 875                         hex_dump(pkt, descs[i].len, descs[i].addr);
 876                 }
 877
 878                 xsk->rx_npkts += rcvd;
 879
 880                 ret = xq_enq(&xsk->tx, descs, rcvd);
 881                 lassert(ret == 0);
 882                 xsk->outstanding_tx += rcvd;
 883         }
 884 }
 885
 886 int main(int argc, char **argv)
 887 {
 888         struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
 889         char xdp_filename[256];
 890         int i, ret, key = 0;
 891         pthread_t pt;
 892
 893         parse_command_line(argc, argv);
 894
 895         if (setrlimit(RLIMIT_MEMLOCK, &r)) {
 896                 fprintf(stderr, "ERROR: setrlimit(RLIMIT_MEMLOCK) \"%s\"\n",
 897                         strerror(errno));
 898                 exit(EXIT_FAILURE);
 899         }
 900
 901         snprintf(xdp_filename, sizeof(xdp_filename), "%s_kern.o", argv[0]);
 902
 903         if (load_bpf_file(xdp_filename)) {
 904                 fprintf(stderr, "ERROR: load_bpf_file %s\n", bpf_log_buf);
 905                 exit(EXIT_FAILURE);
 906         }
 907
 908         if (!prog_fd[0]) {
 909                 fprintf(stderr, "ERROR: load_bpf_file: \"%s\"\n",
 910                         strerror(errno));
 911                 exit(EXIT_FAILURE);
 912         }
 913
 914         if (bpf_set_link_xdp_fd(opt_ifindex, prog_fd[0], opt_xdp_flags) < 0) {
 915                 fprintf(stderr, "ERROR: link set xdp fd failed\n");
 916                 exit(EXIT_FAILURE);
 917         }
 918
 919         ret = bpf_map_update_elem(map_fd[0], &key, &opt_queue, 0);
 920         if (ret) {
 921                 fprintf(stderr, "ERROR: bpf_map_update_elem qidconf\n");
 922                 exit(EXIT_FAILURE);
 923         }
 924
 925         /* Create sockets... */
 926         xsks[num_socks++] = xsk_configure(NULL);
 927
 928 #if RR_LB
 929         for (i = 0; i < MAX_SOCKS - 1; i++)
 930                 xsks[num_socks++] = xsk_configure(xsks[0]->umem);
 931 #endif
 932
 933         /* ...and insert them into the map. */
 934         for (i = 0; i < num_socks; i++) {
 935                 key = i;
 936                 ret = bpf_map_update_elem(map_fd[1], &key, &xsks[i]->sfd, 0);
 937                 if (ret) {
 938                         fprintf(stderr, "ERROR: bpf_map_update_elem %d\n", i);
 939                         exit(EXIT_FAILURE);
 940                 }
 941         }
 942
 943         signal(SIGINT, int_exit);
 944         signal(SIGTERM, int_exit);
 945         signal(SIGABRT, int_exit);
 946
 947         setlocale(LC_ALL, "");
 948
 949         ret = pthread_create(&pt, NULL, poller, NULL);
 950         lassert(ret == 0);
 951
 952         prev_time = get_nsecs();
 953
 954         if (opt_bench == BENCH_RXDROP)
 955                 rx_drop_all();
 956         else if (opt_bench == BENCH_TXONLY)
 957                 tx_only(xsks[0]);
 958         else
 959                 l2fwd(xsks[0]);
 960
 961         return 0;
 962 }