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/libbpf.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 = XDP_FLAGS_UPDATE_IF_NOEXIST;
  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 static __u32 prog_id;
  80
  81 struct xdp_umem_uqueue {
  82         u32 cached_prod;
  83         u32 cached_cons;
  84         u32 mask;
  85         u32 size;
  86         u32 *producer;
  87         u32 *consumer;
  88         u64 *ring;
  89         void *map;
  90 };
  91
  92 struct xdp_umem {
  93         char *frames;
  94         struct xdp_umem_uqueue fq;
  95         struct xdp_umem_uqueue cq;
  96         int fd;
  97 };
  98
  99 struct xdp_uqueue {
 100         u32 cached_prod;
 101         u32 cached_cons;
 102         u32 mask;
 103         u32 size;
 104         u32 *producer;
 105         u32 *consumer;
 106         struct xdp_desc *ring;
 107         void *map;
 108 };
 109
 110 struct xdpsock {
 111         struct xdp_uqueue rx;
 112         struct xdp_uqueue tx;
 113         int sfd;
 114         struct xdp_umem *umem;
 115         u32 outstanding_tx;
 116         unsigned long rx_npkts;
 117         unsigned long tx_npkts;
 118         unsigned long prev_rx_npkts;
 119         unsigned long prev_tx_npkts;
 120 };
 121
 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 #ifdef __aarch64__
 149 #define u_smp_rmb() __asm__ __volatile__("dmb ishld": : :"memory")
 150 #define u_smp_wmb() __asm__ __volatile__("dmb ishst": : :"memory")
 151 #else
 152 #define u_smp_rmb() barrier()
 153 #define u_smp_wmb() barrier()
 154 #endif
 155 #define likely(x) __builtin_expect(!!(x), 1)
 156 #define unlikely(x) __builtin_expect(!!(x), 0)
 157
 158 static const char pkt_data[] =
 159         "\x3c\xfd\xfe\x9e\x7f\x71\xec\xb1\xd7\x98\x3a\xc0\x08\x00\x45\x00"
 160         "\x00\x2e\x00\x00\x00\x00\x40\x11\x88\x97\x05\x08\x07\x08\xc8\x14"
 161         "\x1e\x04\x10\x92\x10\x92\x00\x1a\x6d\xa3\x34\x33\x1f\x69\x40\x6b"
 162         "\x54\x59\xb6\x14\x2d\x11\x44\xbf\xaf\xd9\xbe\xaa";
 163
 164 static inline u32 umem_nb_free(struct xdp_umem_uqueue *q, u32 nb)
 165 {
 166         u32 free_entries = q->cached_cons - q->cached_prod;
 167
 168         if (free_entries >= nb)
 169                 return free_entries;
 170
 171         /* Refresh the local tail pointer */
 172         q->cached_cons = *q->consumer + q->size;
 173
 174         return q->cached_cons - q->cached_prod;
 175 }
 176
 177 static inline u32 xq_nb_free(struct xdp_uqueue *q, u32 ndescs)
 178 {
 179         u32 free_entries = q->cached_cons - q->cached_prod;
 180
 181         if (free_entries >= ndescs)
 182                 return free_entries;
 183
 184         /* Refresh the local tail pointer */
 185         q->cached_cons = *q->consumer + q->size;
 186         return q->cached_cons - q->cached_prod;
 187 }
 188
 189 static inline u32 umem_nb_avail(struct xdp_umem_uqueue *q, u32 nb)
 190 {
 191         u32 entries = q->cached_prod - q->cached_cons;
 192
 193         if (entries == 0) {
 194                 q->cached_prod = *q->producer;
 195                 entries = q->cached_prod - q->cached_cons;
 196         }
 197
 198         return (entries > nb) ? nb : entries;
 199 }
 200
 201 static inline u32 xq_nb_avail(struct xdp_uqueue *q, u32 ndescs)
 202 {
 203         u32 entries = q->cached_prod - q->cached_cons;
 204
 205         if (entries == 0) {
 206                 q->cached_prod = *q->producer;
 207                 entries = q->cached_prod - q->cached_cons;
 208         }
 209
 210         return (entries > ndescs) ? ndescs : entries;
 211 }
 212
 213 static inline int umem_fill_to_kernel_ex(struct xdp_umem_uqueue *fq,
 214                                          struct xdp_desc *d,
 215                                          size_t nb)
 216 {
 217         u32 i;
 218
 219         if (umem_nb_free(fq, nb) < nb)
 220                 return -ENOSPC;
 221
 222         for (i = 0; i < nb; i++) {
 223                 u32 idx = fq->cached_prod++ & fq->mask;
 224
 225                 fq->ring[idx] = d[i].addr;
 226         }
 227
 228         u_smp_wmb();
 229
 230         *fq->producer = fq->cached_prod;
 231
 232         return 0;
 233 }
 234
 235 static inline int umem_fill_to_kernel(struct xdp_umem_uqueue *fq, u64 *d,
 236                                       size_t nb)
 237 {
 238         u32 i;
 239
 240         if (umem_nb_free(fq, nb) < nb)
 241                 return -ENOSPC;
 242
 243         for (i = 0; i < nb; i++) {
 244                 u32 idx = fq->cached_prod++ & fq->mask;
 245
 246                 fq->ring[idx] = d[i];
 247         }
 248
 249         u_smp_wmb();
 250
 251         *fq->producer = fq->cached_prod;
 252
 253         return 0;
 254 }
 255
 256 static inline size_t umem_complete_from_kernel(struct xdp_umem_uqueue *cq,
 257                                                u64 *d, size_t nb)
 258 {
 259         u32 idx, i, entries = umem_nb_avail(cq, nb);
 260
 261         u_smp_rmb();
 262
 263         for (i = 0; i < entries; i++) {
 264                 idx = cq->cached_cons++ & cq->mask;
 265                 d[i] = cq->ring[idx];
 266         }
 267
 268         if (entries > 0) {
 269                 u_smp_wmb();
 270
 271                 *cq->consumer = cq->cached_cons;
 272         }
 273
 274         return entries;
 275 }
 276
 277 static inline void *xq_get_data(struct xdpsock *xsk, u64 addr)
 278 {
 279         return &xsk->umem->frames[addr];
 280 }
 281
 282 static inline int xq_enq(struct xdp_uqueue *uq,
 283                          const struct xdp_desc *descs,
 284                          unsigned int ndescs)
 285 {
 286         struct xdp_desc *r = uq->ring;
 287         unsigned int i;
 288
 289         if (xq_nb_free(uq, ndescs) < ndescs)
 290                 return -ENOSPC;
 291
 292         for (i = 0; i < ndescs; i++) {
 293                 u32 idx = uq->cached_prod++ & uq->mask;
 294
 295                 r[idx].addr = descs[i].addr;
 296                 r[idx].len = descs[i].len;
 297         }
 298
 299         u_smp_wmb();
 300
 301         *uq->producer = uq->cached_prod;
 302         return 0;
 303 }
 304
 305 static inline int xq_enq_tx_only(struct xdp_uqueue *uq,
 306                                  unsigned int id, unsigned int ndescs)
 307 {
 308         struct xdp_desc *r = uq->ring;
 309         unsigned int i;
 310
 311         if (xq_nb_free(uq, ndescs) < ndescs)
 312                 return -ENOSPC;
 313
 314         for (i = 0; i < ndescs; i++) {
 315                 u32 idx = uq->cached_prod++ & uq->mask;
 316
 317                 r[idx].addr     = (id + i) << FRAME_SHIFT;
 318                 r[idx].len      = sizeof(pkt_data) - 1;
 319         }
 320
 321         u_smp_wmb();
 322
 323         *uq->producer = uq->cached_prod;
 324         return 0;
 325 }
 326
 327 static inline int xq_deq(struct xdp_uqueue *uq,
 328                          struct xdp_desc *descs,
 329                          int ndescs)
 330 {
 331         struct xdp_desc *r = uq->ring;
 332         unsigned int idx;
 333         int i, entries;
 334
 335         entries = xq_nb_avail(uq, ndescs);
 336
 337         u_smp_rmb();
 338
 339         for (i = 0; i < entries; i++) {
 340                 idx = uq->cached_cons++ & uq->mask;
 341                 descs[i] = r[idx];
 342         }
 343
 344         if (entries > 0) {
 345                 u_smp_wmb();
 346
 347                 *uq->consumer = uq->cached_cons;
 348         }
 349
 350         return entries;
 351 }
 352
 353 static void swap_mac_addresses(void *data)
 354 {
 355         struct ether_header *eth = (struct ether_header *)data;
 356         struct ether_addr *src_addr = (struct ether_addr *)&eth->ether_shost;
 357         struct ether_addr *dst_addr = (struct ether_addr *)&eth->ether_dhost;
 358         struct ether_addr tmp;
 359
 360         tmp = *src_addr;
 361         *src_addr = *dst_addr;
 362         *dst_addr = tmp;
 363 }
 364
 365 static void hex_dump(void *pkt, size_t length, u64 addr)
 366 {
 367         const unsigned char *address = (unsigned char *)pkt;
 368         const unsigned char *line = address;
 369         size_t line_size = 32;
 370         unsigned char c;
 371         char buf[32];
 372         int i = 0;
 373
 374         if (!DEBUG_HEXDUMP)
 375                 return;
 376
 377         sprintf(buf, "addr=%llu", addr);
 378         printf("length = %zu\n", length);
 379         printf("%s | ", buf);
 380         while (length-- > 0) {
 381                 printf("%02X ", *address++);
 382                 if (!(++i % line_size) || (length == 0 && i % line_size)) {
 383                         if (length == 0) {
 384                                 while (i++ % line_size)
 385                                         printf("__ ");
 386                         }
 387                         printf(" | ");  /* right close */
 388                         while (line < address) {
 389                                 c = *line++;
 390                                 printf("%c", (c < 33 || c == 255) ? 0x2E : c);
 391                         }
 392                         printf("\n");
 393                         if (length > 0)
 394                                 printf("%s | ", buf);
 395                 }
 396         }
 397         printf("\n");
 398 }
 399
 400 static size_t gen_eth_frame(char *frame)
 401 {
 402         memcpy(frame, pkt_data, sizeof(pkt_data) - 1);
 403         return sizeof(pkt_data) - 1;
 404 }
 405
 406 static struct xdp_umem *xdp_umem_configure(int sfd)
 407 {
 408         int fq_size = FQ_NUM_DESCS, cq_size = CQ_NUM_DESCS;
 409         struct xdp_mmap_offsets off;
 410         struct xdp_umem_reg mr;
 411         struct xdp_umem *umem;
 412         socklen_t optlen;
 413         void *bufs;
 414
 415         umem = calloc(1, sizeof(*umem));
 416         lassert(umem);
 417
 418         lassert(posix_memalign(&bufs, getpagesize(), /* PAGE_SIZE aligned */
 419                                NUM_FRAMES * FRAME_SIZE) == 0);
 420
 421         mr.addr = (__u64)bufs;
 422         mr.len = NUM_FRAMES * FRAME_SIZE;
 423         mr.chunk_size = FRAME_SIZE;
 424         mr.headroom = FRAME_HEADROOM;
 425
 426         lassert(setsockopt(sfd, SOL_XDP, XDP_UMEM_REG, &mr, sizeof(mr)) == 0);
 427         lassert(setsockopt(sfd, SOL_XDP, XDP_UMEM_FILL_RING, &fq_size,
 428                            sizeof(int)) == 0);
 429         lassert(setsockopt(sfd, SOL_XDP, XDP_UMEM_COMPLETION_RING, &cq_size,
 430                            sizeof(int)) == 0);
 431
 432         optlen = sizeof(off);
 433         lassert(getsockopt(sfd, SOL_XDP, XDP_MMAP_OFFSETS, &off,
 434                            &optlen) == 0);
 435
 436         umem->fq.map = mmap(0, off.fr.desc +
 437                             FQ_NUM_DESCS * sizeof(u64),
 438                             PROT_READ | PROT_WRITE,
 439                             MAP_SHARED | MAP_POPULATE, sfd,
 440                             XDP_UMEM_PGOFF_FILL_RING);
 441         lassert(umem->fq.map != MAP_FAILED);
 442
 443         umem->fq.mask = FQ_NUM_DESCS - 1;
 444         umem->fq.size = FQ_NUM_DESCS;
 445         umem->fq.producer = umem->fq.map + off.fr.producer;
 446         umem->fq.consumer = umem->fq.map + off.fr.consumer;
 447         umem->fq.ring = umem->fq.map + off.fr.desc;
 448         umem->fq.cached_cons = FQ_NUM_DESCS;
 449
 450         umem->cq.map = mmap(0, off.cr.desc +
 451                              CQ_NUM_DESCS * sizeof(u64),
 452                              PROT_READ | PROT_WRITE,
 453                              MAP_SHARED | MAP_POPULATE, sfd,
 454                              XDP_UMEM_PGOFF_COMPLETION_RING);
 455         lassert(umem->cq.map != MAP_FAILED);
 456
 457         umem->cq.mask = CQ_NUM_DESCS - 1;
 458         umem->cq.size = CQ_NUM_DESCS;
 459         umem->cq.producer = umem->cq.map + off.cr.producer;
 460         umem->cq.consumer = umem->cq.map + off.cr.consumer;
 461         umem->cq.ring = umem->cq.map + off.cr.desc;
 462
 463         umem->frames = bufs;
 464         umem->fd = sfd;
 465
 466         if (opt_bench == BENCH_TXONLY) {
 467                 int i;
 468
 469                 for (i = 0; i < NUM_FRAMES * FRAME_SIZE; i += FRAME_SIZE)
 470                         (void)gen_eth_frame(&umem->frames[i]);
 471         }
 472
 473         return umem;
 474 }
 475
 476 static struct xdpsock *xsk_configure(struct xdp_umem *umem)
 477 {
 478         struct sockaddr_xdp sxdp = {};
 479         struct xdp_mmap_offsets off;
 480         int sfd, ndescs = NUM_DESCS;
 481         struct xdpsock *xsk;
 482         bool shared = true;
 483         socklen_t optlen;
 484         u64 i;
 485
 486         sfd = socket(PF_XDP, SOCK_RAW, 0);
 487         lassert(sfd >= 0);
 488
 489         xsk = calloc(1, sizeof(*xsk));
 490         lassert(xsk);
 491
 492         xsk->sfd = sfd;
 493         xsk->outstanding_tx = 0;
 494
 495         if (!umem) {
 496                 shared = false;
 497                 xsk->umem = xdp_umem_configure(sfd);
 498         } else {
 499                 xsk->umem = umem;
 500         }
 501
 502         lassert(setsockopt(sfd, SOL_XDP, XDP_RX_RING,
 503                            &ndescs, sizeof(int)) == 0);
 504         lassert(setsockopt(sfd, SOL_XDP, XDP_TX_RING,
 505                            &ndescs, sizeof(int)) == 0);
 506         optlen = sizeof(off);
 507         lassert(getsockopt(sfd, SOL_XDP, XDP_MMAP_OFFSETS, &off,
 508                            &optlen) == 0);
 509
 510         /* Rx */
 511         xsk->rx.map = mmap(NULL,
 512                            off.rx.desc +
 513                            NUM_DESCS * sizeof(struct xdp_desc),
 514                            PROT_READ | PROT_WRITE,
 515                            MAP_SHARED | MAP_POPULATE, sfd,
 516                            XDP_PGOFF_RX_RING);
 517         lassert(xsk->rx.map != MAP_FAILED);
 518
 519         if (!shared) {
 520                 for (i = 0; i < NUM_DESCS * FRAME_SIZE; i += FRAME_SIZE)
 521                         lassert(umem_fill_to_kernel(&xsk->umem->fq, &i, 1)
 522                                 == 0);
 523         }
 524
 525         /* Tx */
 526         xsk->tx.map = mmap(NULL,
 527                            off.tx.desc +
 528                            NUM_DESCS * sizeof(struct xdp_desc),
 529                            PROT_READ | PROT_WRITE,
 530                            MAP_SHARED | MAP_POPULATE, sfd,
 531                            XDP_PGOFF_TX_RING);
 532         lassert(xsk->tx.map != MAP_FAILED);
 533
 534         xsk->rx.mask = NUM_DESCS - 1;
 535         xsk->rx.size = NUM_DESCS;
 536         xsk->rx.producer = xsk->rx.map + off.rx.producer;
 537         xsk->rx.consumer = xsk->rx.map + off.rx.consumer;
 538         xsk->rx.ring = xsk->rx.map + off.rx.desc;
 539
 540         xsk->tx.mask = NUM_DESCS - 1;
 541         xsk->tx.size = NUM_DESCS;
 542         xsk->tx.producer = xsk->tx.map + off.tx.producer;
 543         xsk->tx.consumer = xsk->tx.map + off.tx.consumer;
 544         xsk->tx.ring = xsk->tx.map + off.tx.desc;
 545         xsk->tx.cached_cons = NUM_DESCS;
 546
 547         sxdp.sxdp_family = PF_XDP;
 548         sxdp.sxdp_ifindex = opt_ifindex;
 549         sxdp.sxdp_queue_id = opt_queue;
 550
 551         if (shared) {
 552                 sxdp.sxdp_flags = XDP_SHARED_UMEM;
 553                 sxdp.sxdp_shared_umem_fd = umem->fd;
 554         } else {
 555                 sxdp.sxdp_flags = opt_xdp_bind_flags;
 556         }
 557
 558         lassert(bind(sfd, (struct sockaddr *)&sxdp, sizeof(sxdp)) == 0);
 559
 560         return xsk;
 561 }
 562
 563 static void print_benchmark(bool running)
 564 {
 565         const char *bench_str = "INVALID";
 566
 567         if (opt_bench == BENCH_RXDROP)
 568                 bench_str = "rxdrop";
 569         else if (opt_bench == BENCH_TXONLY)
 570                 bench_str = "txonly";
 571         else if (opt_bench == BENCH_L2FWD)
 572                 bench_str = "l2fwd";
 573
 574         printf("%s:%d %s ", opt_if, opt_queue, bench_str);
 575         if (opt_xdp_flags & XDP_FLAGS_SKB_MODE)
 576                 printf("xdp-skb ");
 577         else if (opt_xdp_flags & XDP_FLAGS_DRV_MODE)
 578                 printf("xdp-drv ");
 579         else
 580                 printf("        ");
 581
 582         if (opt_poll)
 583                 printf("poll() ");
 584
 585         if (running) {
 586                 printf("running...");
 587                 fflush(stdout);
 588         }
 589 }
 590
 591 static void dump_stats(void)
 592 {
 593         unsigned long now = get_nsecs();
 594         long dt = now - prev_time;
 595         int i;
 596
 597         prev_time = now;
 598
 599         for (i = 0; i < num_socks && xsks[i]; i++) {
 600                 char *fmt = "%-15s %'-11.0f %'-11lu\n";
 601                 double rx_pps, tx_pps;
 602
 603                 rx_pps = (xsks[i]->rx_npkts - xsks[i]->prev_rx_npkts) *
 604                          1000000000. / dt;
 605                 tx_pps = (xsks[i]->tx_npkts - xsks[i]->prev_tx_npkts) *
 606                          1000000000. / dt;
 607
 608                 printf("\n sock%d@", i);
 609                 print_benchmark(false);
 610                 printf("\n");
 611
 612                 printf("%-15s %-11s %-11s %-11.2f\n", "", "pps", "pkts",
 613                        dt / 1000000000.);
 614                 printf(fmt, "rx", rx_pps, xsks[i]->rx_npkts);
 615                 printf(fmt, "tx", tx_pps, xsks[i]->tx_npkts);
 616
 617                 xsks[i]->prev_rx_npkts = xsks[i]->rx_npkts;
 618                 xsks[i]->prev_tx_npkts = xsks[i]->tx_npkts;
 619         }
 620 }
 621
 622 static void *poller(void *arg)
 623 {
 624         (void)arg;
 625         for (;;) {
 626                 sleep(opt_interval);
 627                 dump_stats();
 628         }
 629
 630         return NULL;
 631 }
 632
 633 static void int_exit(int sig)
 634 {
 635         __u32 curr_prog_id = 0;
 636
 637         (void)sig;
 638         dump_stats();
 639         if (bpf_get_link_xdp_id(opt_ifindex, &curr_prog_id, opt_xdp_flags)) {
 640                 printf("bpf_get_link_xdp_id failed\n");
 641                 exit(EXIT_FAILURE);
 642         }
 643         if (prog_id == curr_prog_id)
 644                 bpf_set_link_xdp_fd(opt_ifindex, -1, opt_xdp_flags);
 645         else if (!curr_prog_id)
 646                 printf("couldn't find a prog id on a given interface\n");
 647         else
 648                 printf("program on interface changed, not removing\n");
 649         exit(EXIT_SUCCESS);
 650 }
 651
 652 static struct option long_options[] = {
 653         {"rxdrop", no_argument, 0, 'r'},
 654         {"txonly", no_argument, 0, 't'},
 655         {"l2fwd", no_argument, 0, 'l'},
 656         {"interface", required_argument, 0, 'i'},
 657         {"queue", required_argument, 0, 'q'},
 658         {"poll", no_argument, 0, 'p'},
 659         {"shared-buffer", no_argument, 0, 's'},
 660         {"xdp-skb", no_argument, 0, 'S'},
 661         {"xdp-native", no_argument, 0, 'N'},
 662         {"interval", required_argument, 0, 'n'},
 663         {"zero-copy", no_argument, 0, 'z'},
 664         {"copy", no_argument, 0, 'c'},
 665         {0, 0, 0, 0}
 666 };
 667
 668 static void usage(const char *prog)
 669 {
 670         const char *str =
 671                 "  Usage: %s [OPTIONS]\n"
 672                 "  Options:\n"
 673                 "  -r, --rxdrop         Discard all incoming packets (default)\n"
 674                 "  -t, --txonly         Only send packets\n"
 675                 "  -l, --l2fwd          MAC swap L2 forwarding\n"
 676                 "  -i, --interface=n    Run on interface n\n"
 677                 "  -q, --queue=n        Use queue n (default 0)\n"
 678                 "  -p, --poll           Use poll syscall\n"
 679                 "  -s, --shared-buffer  Use shared packet buffer\n"
 680                 "  -S, --xdp-skb=n      Use XDP skb-mod\n"
 681                 "  -N, --xdp-native=n   Enfore XDP native mode\n"
 682                 "  -n, --interval=n     Specify statistics update interval (default 1 sec).\n"
 683                 "  -z, --zero-copy      Force zero-copy mode.\n"
 684                 "  -c, --copy           Force copy mode.\n"
 685                 "\n";
 686         fprintf(stderr, str, prog);
 687         exit(EXIT_FAILURE);
 688 }
 689
 690 static void parse_command_line(int argc, char **argv)
 691 {
 692         int option_index, c;
 693
 694         opterr = 0;
 695
 696         for (;;) {
 697                 c = getopt_long(argc, argv, "Frtli:q:psSNn:cz", long_options,
 698                                 &option_index);
 699                 if (c == -1)
 700                         break;
 701
 702                 switch (c) {
 703                 case 'r':
 704                         opt_bench = BENCH_RXDROP;
 705                         break;
 706                 case 't':
 707                         opt_bench = BENCH_TXONLY;
 708                         break;
 709                 case 'l':
 710                         opt_bench = BENCH_L2FWD;
 711                         break;
 712                 case 'i':
 713                         opt_if = optarg;
 714                         break;
 715                 case 'q':
 716                         opt_queue = atoi(optarg);
 717                         break;
 718                 case 's':
 719                         opt_shared_packet_buffer = 1;
 720                         break;
 721                 case 'p':
 722                         opt_poll = 1;
 723                         break;
 724                 case 'S':
 725                         opt_xdp_flags |= XDP_FLAGS_SKB_MODE;
 726                         opt_xdp_bind_flags |= XDP_COPY;
 727                         break;
 728                 case 'N':
 729                         opt_xdp_flags |= XDP_FLAGS_DRV_MODE;
 730                         break;
 731                 case 'n':
 732                         opt_interval = atoi(optarg);
 733                         break;
 734                 case 'z':
 735                         opt_xdp_bind_flags |= XDP_ZEROCOPY;
 736                         break;
 737                 case 'c':
 738                         opt_xdp_bind_flags |= XDP_COPY;
 739                         break;
 740                 case 'F':
 741                         opt_xdp_flags &= ~XDP_FLAGS_UPDATE_IF_NOEXIST;
 742                         break;
 743                 default:
 744                         usage(basename(argv[0]));
 745                 }
 746         }
 747
 748         opt_ifindex = if_nametoindex(opt_if);
 749         if (!opt_ifindex) {
 750                 fprintf(stderr, "ERROR: interface \"%s\" does not exist\n",
 751                         opt_if);
 752                 usage(basename(argv[0]));
 753         }
 754 }
 755
 756 static void kick_tx(int fd)
 757 {
 758         int ret;
 759
 760         ret = sendto(fd, NULL, 0, MSG_DONTWAIT, NULL, 0);
 761         if (ret >= 0 || errno == ENOBUFS || errno == EAGAIN || errno == EBUSY)
 762                 return;
 763         lassert(0);
 764 }
 765
 766 static inline void complete_tx_l2fwd(struct xdpsock *xsk)
 767 {
 768         u64 descs[BATCH_SIZE];
 769         unsigned int rcvd;
 770         size_t ndescs;
 771
 772         if (!xsk->outstanding_tx)
 773                 return;
 774
 775         kick_tx(xsk->sfd);
 776         ndescs = (xsk->outstanding_tx > BATCH_SIZE) ? BATCH_SIZE :
 777                  xsk->outstanding_tx;
 778
 779         /* re-add completed Tx buffers */
 780         rcvd = umem_complete_from_kernel(&xsk->umem->cq, descs, ndescs);
 781         if (rcvd > 0) {
 782                 umem_fill_to_kernel(&xsk->umem->fq, descs, rcvd);
 783                 xsk->outstanding_tx -= rcvd;
 784                 xsk->tx_npkts += rcvd;
 785         }
 786 }
 787
 788 static inline void complete_tx_only(struct xdpsock *xsk)
 789 {
 790         u64 descs[BATCH_SIZE];
 791         unsigned int rcvd;
 792
 793         if (!xsk->outstanding_tx)
 794                 return;
 795
 796         kick_tx(xsk->sfd);
 797
 798         rcvd = umem_complete_from_kernel(&xsk->umem->cq, descs, BATCH_SIZE);
 799         if (rcvd > 0) {
 800                 xsk->outstanding_tx -= rcvd;
 801                 xsk->tx_npkts += rcvd;
 802         }
 803 }
 804
 805 static void rx_drop(struct xdpsock *xsk)
 806 {
 807         struct xdp_desc descs[BATCH_SIZE];
 808         unsigned int rcvd, i;
 809
 810         rcvd = xq_deq(&xsk->rx, descs, BATCH_SIZE);
 811         if (!rcvd)
 812                 return;
 813
 814         for (i = 0; i < rcvd; i++) {
 815                 char *pkt = xq_get_data(xsk, descs[i].addr);
 816
 817                 hex_dump(pkt, descs[i].len, descs[i].addr);
 818         }
 819
 820         xsk->rx_npkts += rcvd;
 821
 822         umem_fill_to_kernel_ex(&xsk->umem->fq, descs, rcvd);
 823 }
 824
 825 static void rx_drop_all(void)
 826 {
 827         struct pollfd fds[MAX_SOCKS + 1];
 828         int i, ret, timeout, nfds = 1;
 829
 830         memset(fds, 0, sizeof(fds));
 831
 832         for (i = 0; i < num_socks; i++) {
 833                 fds[i].fd = xsks[i]->sfd;
 834                 fds[i].events = POLLIN;
 835                 timeout = 1000; /* 1sn */
 836         }
 837
 838         for (;;) {
 839                 if (opt_poll) {
 840                         ret = poll(fds, nfds, timeout);
 841                         if (ret <= 0)
 842                                 continue;
 843                 }
 844
 845                 for (i = 0; i < num_socks; i++)
 846                         rx_drop(xsks[i]);
 847         }
 848 }
 849
 850 static void tx_only(struct xdpsock *xsk)
 851 {
 852         int timeout, ret, nfds = 1;
 853         struct pollfd fds[nfds + 1];
 854         unsigned int idx = 0;
 855
 856         memset(fds, 0, sizeof(fds));
 857         fds[0].fd = xsk->sfd;
 858         fds[0].events = POLLOUT;
 859         timeout = 1000; /* 1sn */
 860
 861         for (;;) {
 862                 if (opt_poll) {
 863                         ret = poll(fds, nfds, timeout);
 864                         if (ret <= 0)
 865                                 continue;
 866
 867                         if (fds[0].fd != xsk->sfd ||
 868                             !(fds[0].revents & POLLOUT))
 869                                 continue;
 870                 }
 871
 872                 if (xq_nb_free(&xsk->tx, BATCH_SIZE) >= BATCH_SIZE) {
 873                         lassert(xq_enq_tx_only(&xsk->tx, idx, BATCH_SIZE) == 0);
 874
 875                         xsk->outstanding_tx += BATCH_SIZE;
 876                         idx += BATCH_SIZE;
 877                         idx %= NUM_FRAMES;
 878                 }
 879
 880                 complete_tx_only(xsk);
 881         }
 882 }
 883
 884 static void l2fwd(struct xdpsock *xsk)
 885 {
 886         for (;;) {
 887                 struct xdp_desc descs[BATCH_SIZE];
 888                 unsigned int rcvd, i;
 889                 int ret;
 890
 891                 for (;;) {
 892                         complete_tx_l2fwd(xsk);
 893
 894                         rcvd = xq_deq(&xsk->rx, descs, BATCH_SIZE);
 895                         if (rcvd > 0)
 896                                 break;
 897                 }
 898
 899                 for (i = 0; i < rcvd; i++) {
 900                         char *pkt = xq_get_data(xsk, descs[i].addr);
 901
 902                         swap_mac_addresses(pkt);
 903
 904                         hex_dump(pkt, descs[i].len, descs[i].addr);
 905                 }
 906
 907                 xsk->rx_npkts += rcvd;
 908
 909                 ret = xq_enq(&xsk->tx, descs, rcvd);
 910                 lassert(ret == 0);
 911                 xsk->outstanding_tx += rcvd;
 912         }
 913 }
 914
 915 int main(int argc, char **argv)
 916 {
 917         struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
 918         struct bpf_prog_load_attr prog_load_attr = {
 919                 .prog_type      = BPF_PROG_TYPE_XDP,
 920         };
 921         int prog_fd, qidconf_map, xsks_map;
 922         struct bpf_prog_info info = {};
 923         __u32 info_len = sizeof(info);
 924         struct bpf_object *obj;
 925         char xdp_filename[256];
 926         struct bpf_map *map;
 927         int i, ret, key = 0;
 928         pthread_t pt;
 929
 930         parse_command_line(argc, argv);
 931
 932         if (setrlimit(RLIMIT_MEMLOCK, &r)) {
 933                 fprintf(stderr, "ERROR: setrlimit(RLIMIT_MEMLOCK) \"%s\"\n",
 934                         strerror(errno));
 935                 exit(EXIT_FAILURE);
 936         }
 937
 938         snprintf(xdp_filename, sizeof(xdp_filename), "%s_kern.o", argv[0]);
 939         prog_load_attr.file = xdp_filename;
 940
 941         if (bpf_prog_load_xattr(&prog_load_attr, &obj, &prog_fd))
 942                 exit(EXIT_FAILURE);
 943         if (prog_fd < 0) {
 944                 fprintf(stderr, "ERROR: no program found: %s\n",
 945                         strerror(prog_fd));
 946                 exit(EXIT_FAILURE);
 947         }
 948
 949         map = bpf_object__find_map_by_name(obj, "qidconf_map");
 950         qidconf_map = bpf_map__fd(map);
 951         if (qidconf_map < 0) {
 952                 fprintf(stderr, "ERROR: no qidconf map found: %s\n",
 953                         strerror(qidconf_map));
 954                 exit(EXIT_FAILURE);
 955         }
 956
 957         map = bpf_object__find_map_by_name(obj, "xsks_map");
 958         xsks_map = bpf_map__fd(map);
 959         if (xsks_map < 0) {
 960                 fprintf(stderr, "ERROR: no xsks map found: %s\n",
 961                         strerror(xsks_map));
 962                 exit(EXIT_FAILURE);
 963         }
 964
 965         if (bpf_set_link_xdp_fd(opt_ifindex, prog_fd, opt_xdp_flags) < 0) {
 966                 fprintf(stderr, "ERROR: link set xdp fd failed\n");
 967                 exit(EXIT_FAILURE);
 968         }
 969
 970         ret = bpf_obj_get_info_by_fd(prog_fd, &info, &info_len);
 971         if (ret) {
 972                 printf("can't get prog info - %s\n", strerror(errno));
 973                 return 1;
 974         }
 975         prog_id = info.id;
 976
 977         ret = bpf_map_update_elem(qidconf_map, &key, &opt_queue, 0);
 978         if (ret) {
 979                 fprintf(stderr, "ERROR: bpf_map_update_elem qidconf\n");
 980                 exit(EXIT_FAILURE);
 981         }
 982
 983         /* Create sockets... */
 984         xsks[num_socks++] = xsk_configure(NULL);
 985
 986 #if RR_LB
 987         for (i = 0; i < MAX_SOCKS - 1; i++)
 988                 xsks[num_socks++] = xsk_configure(xsks[0]->umem);
 989 #endif
 990
 991         /* ...and insert them into the map. */
 992         for (i = 0; i < num_socks; i++) {
 993                 key = i;
 994                 ret = bpf_map_update_elem(xsks_map, &key, &xsks[i]->sfd, 0);
 995                 if (ret) {
 996                         fprintf(stderr, "ERROR: bpf_map_update_elem %d\n", i);
 997                         exit(EXIT_FAILURE);
 998                 }
 999         }
1000
1001         signal(SIGINT, int_exit);
1002         signal(SIGTERM, int_exit);
1003         signal(SIGABRT, int_exit);
1004
1005         setlocale(LC_ALL, "");
1006
1007         ret = pthread_create(&pt, NULL, poller, NULL);
1008         lassert(ret == 0);
1009
1010         prev_time = get_nsecs();
1011
1012         if (opt_bench == BENCH_RXDROP)
1013                 rx_drop_all();
1014         else if (opt_bench == BENCH_TXONLY)
1015                 tx_only(xsks[0]);
1016         else
1017                 l2fwd(xsks[0]);
1018
1019         return 0;
1020 }