1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2017 - 2018 Intel Corporation. */
4 #include <asm/barrier.h>
9 #include <linux/compiler.h>
10 #include <linux/if_link.h>
11 #include <linux/if_xdp.h>
12 #include <linux/if_ether.h>
14 #include <net/ethernet.h>
24 #include <sys/resource.h>
25 #include <sys/socket.h>
26 #include <sys/types.h>
46 #define NUM_FRAMES (4 * 1024)
49 #define DEBUG_HEXDUMP 0
55 static unsigned long prev_time;
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;
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;
80 struct xsk_umem_info {
81 struct xsk_ring_prod fq;
82 struct xsk_ring_cons cq;
83 struct xsk_umem *umem;
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;
100 struct xsk_socket_info *xsks[MAX_SOCKS];
102 static unsigned long get_nsecs(void)
106 clock_gettime(CLOCK_MONOTONIC, &ts);
107 return ts.tv_sec * 1000000000UL + ts.tv_nsec;
110 static void print_benchmark(bool running)
112 const char *bench_str = "INVALID";
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)
121 printf("%s:%d %s ", opt_if, opt_queue, bench_str);
122 if (opt_xdp_flags & XDP_FLAGS_SKB_MODE)
124 else if (opt_xdp_flags & XDP_FLAGS_DRV_MODE)
133 printf("running...");
138 static void dump_stats(void)
140 unsigned long now = get_nsecs();
141 long dt = now - prev_time;
146 for (i = 0; i < num_socks && xsks[i]; i++) {
147 char *fmt = "%-15s %'-11.0f %'-11lu\n";
148 double rx_pps, tx_pps;
150 rx_pps = (xsks[i]->rx_npkts - xsks[i]->prev_rx_npkts) *
152 tx_pps = (xsks[i]->tx_npkts - xsks[i]->prev_tx_npkts) *
155 printf("\n sock%d@", i);
156 print_benchmark(false);
159 printf("%-15s %-11s %-11s %-11.2f\n", "", "pps", "pkts",
161 printf(fmt, "rx", rx_pps, xsks[i]->rx_npkts);
162 printf(fmt, "tx", tx_pps, xsks[i]->tx_npkts);
164 xsks[i]->prev_rx_npkts = xsks[i]->rx_npkts;
165 xsks[i]->prev_tx_npkts = xsks[i]->tx_npkts;
169 static void *poller(void *arg)
180 static void remove_xdp_program(void)
182 __u32 curr_prog_id = 0;
184 if (bpf_get_link_xdp_id(opt_ifindex, &curr_prog_id, opt_xdp_flags)) {
185 printf("bpf_get_link_xdp_id failed\n");
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");
193 printf("program on interface changed, not removing\n");
196 static void int_exit(int sig)
198 struct xsk_umem *umem = xsks[0]->umem->umem;
203 xsk_socket__delete(xsks[0]->xsk);
204 (void)xsk_umem__delete(umem);
205 remove_xdp_program();
210 static void __exit_with_error(int error, const char *file, const char *func,
213 fprintf(stderr, "%s:%s:%i: errno: %d/\"%s\"\n", file, func,
214 line, error, strerror(error));
216 remove_xdp_program();
220 #define exit_with_error(error) __exit_with_error(error, __FILE__, __func__, \
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";
229 static void swap_mac_addresses(void *data)
231 struct ether_header *eth = (struct ether_header *)data;
232 struct ether_addr *src_addr = (struct ether_addr *)ð->ether_shost;
233 struct ether_addr *dst_addr = (struct ether_addr *)ð->ether_dhost;
234 struct ether_addr tmp;
237 *src_addr = *dst_addr;
241 static void hex_dump(void *pkt, size_t length, u64 addr)
243 const unsigned char *address = (unsigned char *)pkt;
244 const unsigned char *line = address;
245 size_t line_size = 32;
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)) {
260 while (i++ % line_size)
263 printf(" | "); /* right close */
264 while (line < address) {
266 printf("%c", (c < 33 || c == 255) ? 0x2E : c);
270 printf("%s | ", buf);
276 static size_t gen_eth_frame(struct xsk_umem_info *umem, u64 addr)
278 memcpy(xsk_umem__get_data(umem->buffer, addr), pkt_data,
279 sizeof(pkt_data) - 1);
280 return sizeof(pkt_data) - 1;
283 static struct xsk_umem_info *xsk_configure_umem(void *buffer, u64 size)
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
296 umem = calloc(1, sizeof(*umem));
298 exit_with_error(errno);
300 ret = xsk_umem__create(&umem->umem, buffer, size, &umem->fq, &umem->cq,
304 exit_with_error(-ret);
306 umem->buffer = buffer;
310 static struct xsk_socket_info *xsk_configure_socket(struct xsk_umem_info *umem)
312 struct xsk_socket_config cfg;
313 struct xsk_socket_info *xsk;
318 xsk = calloc(1, sizeof(*xsk));
320 exit_with_error(errno);
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);
331 exit_with_error(-ret);
333 ret = bpf_get_link_xdp_id(opt_ifindex, &prog_id, opt_xdp_flags);
335 exit_with_error(-ret);
337 ret = xsk_ring_prod__reserve(&xsk->umem->fq,
338 XSK_RING_PROD__DEFAULT_NUM_DESCS,
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);
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'},
369 static void usage(const char *prog)
372 " Usage: %s [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"
390 fprintf(stderr, str, prog, XSK_UMEM__DEFAULT_FRAME_SIZE);
394 static void parse_command_line(int argc, char **argv)
401 c = getopt_long(argc, argv, "Frtli:q:psSNn:czf:mu",
402 long_options, &option_index);
408 opt_bench = BENCH_RXDROP;
411 opt_bench = BENCH_TXONLY;
414 opt_bench = BENCH_L2FWD;
420 opt_queue = atoi(optarg);
426 opt_xdp_flags |= XDP_FLAGS_SKB_MODE;
427 opt_xdp_bind_flags |= XDP_COPY;
430 opt_xdp_flags |= XDP_FLAGS_DRV_MODE;
433 opt_interval = atoi(optarg);
436 opt_xdp_bind_flags |= XDP_ZEROCOPY;
439 opt_xdp_bind_flags |= XDP_COPY;
442 opt_umem_flags |= XDP_UMEM_UNALIGNED_CHUNK_FLAG;
443 opt_unaligned_chunks = 1;
444 opt_mmap_flags = MAP_HUGETLB;
447 opt_xdp_flags &= ~XDP_FLAGS_UPDATE_IF_NOEXIST;
450 opt_xsk_frame_size = atoi(optarg);
452 opt_need_wakeup = false;
453 opt_xdp_bind_flags &= ~XDP_USE_NEED_WAKEUP;
457 usage(basename(argv[0]));
461 opt_ifindex = if_nametoindex(opt_if);
463 fprintf(stderr, "ERROR: interface \"%s\" does not exist\n",
465 usage(basename(argv[0]));
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",
472 usage(basename(argv[0]));
476 static void kick_tx(struct xsk_socket_info *xsk)
480 ret = sendto(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, 0);
481 if (ret >= 0 || errno == ENOBUFS || errno == EAGAIN || errno == EBUSY)
483 exit_with_error(errno);
486 static inline void complete_tx_l2fwd(struct xsk_socket_info *xsk,
489 struct xsk_umem_info *umem = xsk->umem;
490 u32 idx_cq = 0, idx_fq = 0;
494 if (!xsk->outstanding_tx)
497 if (!opt_need_wakeup || xsk_ring_prod__needs_wakeup(&xsk->tx))
500 ndescs = (xsk->outstanding_tx > BATCH_SIZE) ? BATCH_SIZE :
503 /* re-add completed Tx buffers */
504 rcvd = xsk_ring_cons__peek(&umem->cq, ndescs, &idx_cq);
509 ret = xsk_ring_prod__reserve(&umem->fq, rcvd, &idx_fq);
510 while (ret != rcvd) {
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);
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++);
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;
529 static inline void complete_tx_only(struct xsk_socket_info *xsk)
534 if (!xsk->outstanding_tx)
537 if (!opt_need_wakeup || xsk_ring_prod__needs_wakeup(&xsk->tx))
540 rcvd = xsk_ring_cons__peek(&xsk->umem->cq, BATCH_SIZE, &idx);
542 xsk_ring_cons__release(&xsk->umem->cq, rcvd);
543 xsk->outstanding_tx -= rcvd;
544 xsk->tx_npkts += rcvd;
548 static void rx_drop(struct xsk_socket_info *xsk, struct pollfd *fds)
550 unsigned int rcvd, i;
551 u32 idx_rx = 0, idx_fq = 0;
554 rcvd = xsk_ring_cons__peek(&xsk->rx, BATCH_SIZE, &idx_rx);
556 if (xsk_ring_prod__needs_wakeup(&xsk->umem->fq))
557 ret = poll(fds, num_socks, opt_timeout);
561 ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
562 while (ret != rcvd) {
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);
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);
575 addr = xsk_umem__add_offset_to_addr(addr);
576 char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
578 hex_dump(pkt, len, addr);
579 *xsk_ring_prod__fill_addr(&xsk->umem->fq, idx_fq++) = orig;
582 xsk_ring_prod__submit(&xsk->umem->fq, rcvd);
583 xsk_ring_cons__release(&xsk->rx, rcvd);
584 xsk->rx_npkts += rcvd;
587 static void rx_drop_all(void)
589 struct pollfd fds[MAX_SOCKS + 1];
592 memset(fds, 0, sizeof(fds));
594 for (i = 0; i < num_socks; i++) {
595 fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
596 fds[i].events = POLLIN;
601 ret = poll(fds, num_socks, opt_timeout);
606 for (i = 0; i < num_socks; i++)
607 rx_drop(xsks[i], fds);
611 static void tx_only(struct xsk_socket_info *xsk, u32 frame_nb)
615 if (xsk_ring_prod__reserve(&xsk->tx, BATCH_SIZE, &idx) == BATCH_SIZE) {
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;
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;
631 complete_tx_only(xsk);
634 static void tx_only_all(void)
636 struct pollfd fds[MAX_SOCKS];
637 u32 frame_nb[MAX_SOCKS] = {};
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;
648 ret = poll(fds, num_socks, opt_timeout);
652 if (!(fds[0].revents & POLLOUT))
656 for (i = 0; i < num_socks; i++)
657 tx_only(xsks[i], frame_nb[i]);
661 static void l2fwd(struct xsk_socket_info *xsk, struct pollfd *fds)
663 unsigned int rcvd, i;
664 u32 idx_rx = 0, idx_tx = 0;
667 complete_tx_l2fwd(xsk, fds);
669 rcvd = xsk_ring_cons__peek(&xsk->rx, BATCH_SIZE, &idx_rx);
671 if (xsk_ring_prod__needs_wakeup(&xsk->umem->fq))
672 ret = poll(fds, num_socks, opt_timeout);
676 ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
677 while (ret != rcvd) {
679 exit_with_error(-ret);
680 if (xsk_ring_prod__needs_wakeup(&xsk->tx))
682 ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
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 = xsk_umem__extract_addr(addr);
690 addr = xsk_umem__add_offset_to_addr(addr);
691 char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
693 swap_mac_addresses(pkt);
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;
700 xsk_ring_prod__submit(&xsk->tx, rcvd);
701 xsk_ring_cons__release(&xsk->rx, rcvd);
703 xsk->rx_npkts += rcvd;
704 xsk->outstanding_tx += rcvd;
707 static void l2fwd_all(void)
709 struct pollfd fds[MAX_SOCKS];
712 memset(fds, 0, sizeof(fds));
714 for (i = 0; i < num_socks; i++) {
715 fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
716 fds[i].events = POLLOUT | POLLIN;
721 ret = poll(fds, num_socks, opt_timeout);
726 for (i = 0; i < num_socks; i++)
731 int main(int argc, char **argv)
733 struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
734 struct xsk_umem_info *umem;
739 parse_command_line(argc, argv);
741 if (setrlimit(RLIMIT_MEMLOCK, &r)) {
742 fprintf(stderr, "ERROR: setrlimit(RLIMIT_MEMLOCK) \"%s\"\n",
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");
755 /* Create sockets... */
756 umem = xsk_configure_umem(bufs, NUM_FRAMES * opt_xsk_frame_size);
757 xsks[num_socks++] = xsk_configure_socket(umem);
759 if (opt_bench == BENCH_TXONLY) {
762 for (i = 0; i < NUM_FRAMES; i++)
763 (void)gen_eth_frame(umem, i * opt_xsk_frame_size);
766 signal(SIGINT, int_exit);
767 signal(SIGTERM, int_exit);
768 signal(SIGABRT, int_exit);
770 setlocale(LC_ALL, "");
772 ret = pthread_create(&pt, NULL, poller, NULL);
774 exit_with_error(ret);
776 prev_time = get_nsecs();
778 if (opt_bench == BENCH_RXDROP)
780 else if (opt_bench == BENCH_TXONLY)