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 <linux/udp.h>
15 #include <arpa/inet.h>
17 #include <net/ethernet.h>
27 #include <sys/resource.h>
28 #include <sys/socket.h>
29 #include <sys/types.h>
33 #include <bpf/libbpf.h>
50 #define NUM_FRAMES (4 * 1024)
51 #define MIN_PKT_SIZE 64
53 #define DEBUG_HEXDUMP 0
60 static unsigned long prev_time;
68 static enum benchmark_type opt_bench = BENCH_RXDROP;
69 static u32 opt_xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
70 static const char *opt_if = "";
71 static int opt_ifindex;
73 static unsigned long opt_duration;
74 static unsigned long start_time;
75 static bool benchmark_done;
76 static u32 opt_batch_size = 64;
77 static int opt_pkt_count;
78 static u16 opt_pkt_size = MIN_PKT_SIZE;
79 static u32 opt_pkt_fill_pattern = 0x12345678;
81 static int opt_interval = 1;
82 static u32 opt_xdp_bind_flags = XDP_USE_NEED_WAKEUP;
83 static u32 opt_umem_flags;
84 static int opt_unaligned_chunks;
85 static int opt_mmap_flags;
86 static int opt_xsk_frame_size = XSK_UMEM__DEFAULT_FRAME_SIZE;
87 static int opt_timeout = 1000;
88 static bool opt_need_wakeup = true;
89 static u32 opt_num_xsks = 1;
92 struct xsk_umem_info {
93 struct xsk_ring_prod fq;
94 struct xsk_ring_cons cq;
95 struct xsk_umem *umem;
99 struct xsk_socket_info {
100 struct xsk_ring_cons rx;
101 struct xsk_ring_prod tx;
102 struct xsk_umem_info *umem;
103 struct xsk_socket *xsk;
104 unsigned long rx_npkts;
105 unsigned long tx_npkts;
106 unsigned long prev_rx_npkts;
107 unsigned long prev_tx_npkts;
111 static int num_socks;
112 struct xsk_socket_info *xsks[MAX_SOCKS];
114 static unsigned long get_nsecs(void)
118 clock_gettime(CLOCK_MONOTONIC, &ts);
119 return ts.tv_sec * 1000000000UL + ts.tv_nsec;
122 static void print_benchmark(bool running)
124 const char *bench_str = "INVALID";
126 if (opt_bench == BENCH_RXDROP)
127 bench_str = "rxdrop";
128 else if (opt_bench == BENCH_TXONLY)
129 bench_str = "txonly";
130 else if (opt_bench == BENCH_L2FWD)
133 printf("%s:%d %s ", opt_if, opt_queue, bench_str);
134 if (opt_xdp_flags & XDP_FLAGS_SKB_MODE)
136 else if (opt_xdp_flags & XDP_FLAGS_DRV_MODE)
145 printf("running...");
150 static void dump_stats(void)
152 unsigned long now = get_nsecs();
153 long dt = now - prev_time;
158 for (i = 0; i < num_socks && xsks[i]; i++) {
159 char *fmt = "%-15s %'-11.0f %'-11lu\n";
160 double rx_pps, tx_pps;
162 rx_pps = (xsks[i]->rx_npkts - xsks[i]->prev_rx_npkts) *
164 tx_pps = (xsks[i]->tx_npkts - xsks[i]->prev_tx_npkts) *
167 printf("\n sock%d@", i);
168 print_benchmark(false);
171 printf("%-15s %-11s %-11s %-11.2f\n", "", "pps", "pkts",
173 printf(fmt, "rx", rx_pps, xsks[i]->rx_npkts);
174 printf(fmt, "tx", tx_pps, xsks[i]->tx_npkts);
176 xsks[i]->prev_rx_npkts = xsks[i]->rx_npkts;
177 xsks[i]->prev_tx_npkts = xsks[i]->tx_npkts;
181 static bool is_benchmark_done(void)
183 if (opt_duration > 0) {
184 unsigned long dt = (get_nsecs() - start_time);
186 if (dt >= opt_duration)
187 benchmark_done = true;
189 return benchmark_done;
192 static void *poller(void *arg)
195 while (!is_benchmark_done()) {
203 static void remove_xdp_program(void)
205 u32 curr_prog_id = 0;
207 if (bpf_get_link_xdp_id(opt_ifindex, &curr_prog_id, opt_xdp_flags)) {
208 printf("bpf_get_link_xdp_id failed\n");
211 if (prog_id == curr_prog_id)
212 bpf_set_link_xdp_fd(opt_ifindex, -1, opt_xdp_flags);
213 else if (!curr_prog_id)
214 printf("couldn't find a prog id on a given interface\n");
216 printf("program on interface changed, not removing\n");
219 static void int_exit(int sig)
221 benchmark_done = true;
224 static void xdpsock_cleanup(void)
226 struct xsk_umem *umem = xsks[0]->umem->umem;
230 for (i = 0; i < num_socks; i++)
231 xsk_socket__delete(xsks[i]->xsk);
232 (void)xsk_umem__delete(umem);
233 remove_xdp_program();
236 static void __exit_with_error(int error, const char *file, const char *func,
239 fprintf(stderr, "%s:%s:%i: errno: %d/\"%s\"\n", file, func,
240 line, error, strerror(error));
242 remove_xdp_program();
246 #define exit_with_error(error) __exit_with_error(error, __FILE__, __func__, \
248 static void swap_mac_addresses(void *data)
250 struct ether_header *eth = (struct ether_header *)data;
251 struct ether_addr *src_addr = (struct ether_addr *)ð->ether_shost;
252 struct ether_addr *dst_addr = (struct ether_addr *)ð->ether_dhost;
253 struct ether_addr tmp;
256 *src_addr = *dst_addr;
260 static void hex_dump(void *pkt, size_t length, u64 addr)
262 const unsigned char *address = (unsigned char *)pkt;
263 const unsigned char *line = address;
264 size_t line_size = 32;
272 sprintf(buf, "addr=%llu", addr);
273 printf("length = %zu\n", length);
274 printf("%s | ", buf);
275 while (length-- > 0) {
276 printf("%02X ", *address++);
277 if (!(++i % line_size) || (length == 0 && i % line_size)) {
279 while (i++ % line_size)
282 printf(" | "); /* right close */
283 while (line < address) {
285 printf("%c", (c < 33 || c == 255) ? 0x2E : c);
289 printf("%s | ", buf);
295 static void *memset32_htonl(void *dest, u32 val, u32 size)
297 u32 *ptr = (u32 *)dest;
302 for (i = 0; i < (size & (~0x3)); i += 4)
305 for (; i < size; i++)
306 ((char *)dest)[i] = ((char *)&val)[i & 3];
312 * This function code has been taken from
313 * Linux kernel lib/checksum.c
315 static inline unsigned short from32to16(unsigned int x)
317 /* add up 16-bit and 16-bit for 16+c bit */
318 x = (x & 0xffff) + (x >> 16);
320 x = (x & 0xffff) + (x >> 16);
325 * This function code has been taken from
326 * Linux kernel lib/checksum.c
328 static unsigned int do_csum(const unsigned char *buff, int len)
330 unsigned int result = 0;
335 odd = 1 & (unsigned long)buff;
337 #ifdef __LITTLE_ENDIAN
338 result += (*buff << 8);
346 if (2 & (unsigned long)buff) {
347 result += *(unsigned short *)buff;
352 const unsigned char *end = buff +
353 ((unsigned int)len & ~3);
354 unsigned int carry = 0;
357 unsigned int w = *(unsigned int *)buff;
362 carry = (w > result);
363 } while (buff < end);
365 result = (result & 0xffff) + (result >> 16);
368 result += *(unsigned short *)buff;
373 #ifdef __LITTLE_ENDIAN
376 result += (*buff << 8);
378 result = from32to16(result);
380 result = ((result >> 8) & 0xff) | ((result & 0xff) << 8);
385 __sum16 ip_fast_csum(const void *iph, unsigned int ihl);
388 * This is a version of ip_compute_csum() optimized for IP headers,
389 * which always checksum on 4 octet boundaries.
390 * This function code has been taken from
391 * Linux kernel lib/checksum.c
393 __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
395 return (__force __sum16)~do_csum(iph, ihl * 4);
399 * Fold a partial checksum
400 * This function code has been taken from
401 * Linux kernel include/asm-generic/checksum.h
403 static inline __sum16 csum_fold(__wsum csum)
405 u32 sum = (__force u32)csum;
407 sum = (sum & 0xffff) + (sum >> 16);
408 sum = (sum & 0xffff) + (sum >> 16);
409 return (__force __sum16)~sum;
413 * This function code has been taken from
414 * Linux kernel lib/checksum.c
416 static inline u32 from64to32(u64 x)
418 /* add up 32-bit and 32-bit for 32+c bit */
419 x = (x & 0xffffffff) + (x >> 32);
421 x = (x & 0xffffffff) + (x >> 32);
425 __wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
426 __u32 len, __u8 proto, __wsum sum);
429 * This function code has been taken from
430 * Linux kernel lib/checksum.c
432 __wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
433 __u32 len, __u8 proto, __wsum sum)
435 unsigned long long s = (__force u32)sum;
437 s += (__force u32)saddr;
438 s += (__force u32)daddr;
439 #ifdef __BIG_ENDIAN__
442 s += (proto + len) << 8;
444 return (__force __wsum)from64to32(s);
448 * This function has been taken from
449 * Linux kernel include/asm-generic/checksum.h
451 static inline __sum16
452 csum_tcpudp_magic(__be32 saddr, __be32 daddr, __u32 len,
453 __u8 proto, __wsum sum)
455 return csum_fold(csum_tcpudp_nofold(saddr, daddr, len, proto, sum));
458 static inline u16 udp_csum(u32 saddr, u32 daddr, u32 len,
459 u8 proto, u16 *udp_pkt)
464 /* udp hdr and data */
465 for (; cnt < len; cnt += 2)
466 csum += udp_pkt[cnt >> 1];
468 return csum_tcpudp_magic(saddr, daddr, len, proto, csum);
471 #define ETH_FCS_SIZE 4
473 #define PKT_HDR_SIZE (sizeof(struct ethhdr) + sizeof(struct iphdr) + \
474 sizeof(struct udphdr))
476 #define PKT_SIZE (opt_pkt_size - ETH_FCS_SIZE)
477 #define IP_PKT_SIZE (PKT_SIZE - sizeof(struct ethhdr))
478 #define UDP_PKT_SIZE (IP_PKT_SIZE - sizeof(struct iphdr))
479 #define UDP_PKT_DATA_SIZE (UDP_PKT_SIZE - sizeof(struct udphdr))
481 static u8 pkt_data[XSK_UMEM__DEFAULT_FRAME_SIZE];
483 static void gen_eth_hdr_data(void)
485 struct udphdr *udp_hdr = (struct udphdr *)(pkt_data +
486 sizeof(struct ethhdr) +
487 sizeof(struct iphdr));
488 struct iphdr *ip_hdr = (struct iphdr *)(pkt_data +
489 sizeof(struct ethhdr));
490 struct ethhdr *eth_hdr = (struct ethhdr *)pkt_data;
492 /* ethernet header */
493 memcpy(eth_hdr->h_dest, "\x3c\xfd\xfe\x9e\x7f\x71", ETH_ALEN);
494 memcpy(eth_hdr->h_source, "\xec\xb1\xd7\x98\x3a\xc0", ETH_ALEN);
495 eth_hdr->h_proto = htons(ETH_P_IP);
498 ip_hdr->version = IPVERSION;
499 ip_hdr->ihl = 0x5; /* 20 byte header */
501 ip_hdr->tot_len = htons(IP_PKT_SIZE);
503 ip_hdr->frag_off = 0;
504 ip_hdr->ttl = IPDEFTTL;
505 ip_hdr->protocol = IPPROTO_UDP;
506 ip_hdr->saddr = htonl(0x0a0a0a10);
507 ip_hdr->daddr = htonl(0x0a0a0a20);
509 /* IP header checksum */
511 ip_hdr->check = ip_fast_csum((const void *)ip_hdr, ip_hdr->ihl);
514 udp_hdr->source = htons(0x1000);
515 udp_hdr->dest = htons(0x1000);
516 udp_hdr->len = htons(UDP_PKT_SIZE);
519 memset32_htonl(pkt_data + PKT_HDR_SIZE, opt_pkt_fill_pattern,
522 /* UDP header checksum */
524 udp_hdr->check = udp_csum(ip_hdr->saddr, ip_hdr->daddr, UDP_PKT_SIZE,
525 IPPROTO_UDP, (u16 *)udp_hdr);
528 static void gen_eth_frame(struct xsk_umem_info *umem, u64 addr)
530 memcpy(xsk_umem__get_data(umem->buffer, addr), pkt_data,
534 static struct xsk_umem_info *xsk_configure_umem(void *buffer, u64 size)
536 struct xsk_umem_info *umem;
537 struct xsk_umem_config cfg = {
538 .fill_size = XSK_RING_PROD__DEFAULT_NUM_DESCS,
539 .comp_size = XSK_RING_CONS__DEFAULT_NUM_DESCS,
540 .frame_size = opt_xsk_frame_size,
541 .frame_headroom = XSK_UMEM__DEFAULT_FRAME_HEADROOM,
542 .flags = opt_umem_flags
546 umem = calloc(1, sizeof(*umem));
548 exit_with_error(errno);
550 ret = xsk_umem__create(&umem->umem, buffer, size, &umem->fq, &umem->cq,
553 exit_with_error(-ret);
555 umem->buffer = buffer;
559 static void xsk_populate_fill_ring(struct xsk_umem_info *umem)
564 ret = xsk_ring_prod__reserve(&umem->fq,
565 XSK_RING_PROD__DEFAULT_NUM_DESCS, &idx);
566 if (ret != XSK_RING_PROD__DEFAULT_NUM_DESCS)
567 exit_with_error(-ret);
568 for (i = 0; i < XSK_RING_PROD__DEFAULT_NUM_DESCS; i++)
569 *xsk_ring_prod__fill_addr(&umem->fq, idx++) =
570 i * opt_xsk_frame_size;
571 xsk_ring_prod__submit(&umem->fq, XSK_RING_PROD__DEFAULT_NUM_DESCS);
574 static struct xsk_socket_info *xsk_configure_socket(struct xsk_umem_info *umem,
577 struct xsk_socket_config cfg;
578 struct xsk_socket_info *xsk;
579 struct xsk_ring_cons *rxr;
580 struct xsk_ring_prod *txr;
583 xsk = calloc(1, sizeof(*xsk));
585 exit_with_error(errno);
588 cfg.rx_size = XSK_RING_CONS__DEFAULT_NUM_DESCS;
589 cfg.tx_size = XSK_RING_PROD__DEFAULT_NUM_DESCS;
590 if (opt_num_xsks > 1)
591 cfg.libbpf_flags = XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD;
593 cfg.libbpf_flags = 0;
594 cfg.xdp_flags = opt_xdp_flags;
595 cfg.bind_flags = opt_xdp_bind_flags;
597 rxr = rx ? &xsk->rx : NULL;
598 txr = tx ? &xsk->tx : NULL;
599 ret = xsk_socket__create(&xsk->xsk, opt_if, opt_queue, umem->umem,
602 exit_with_error(-ret);
604 ret = bpf_get_link_xdp_id(opt_ifindex, &prog_id, opt_xdp_flags);
606 exit_with_error(-ret);
611 static struct option long_options[] = {
612 {"rxdrop", no_argument, 0, 'r'},
613 {"txonly", no_argument, 0, 't'},
614 {"l2fwd", no_argument, 0, 'l'},
615 {"interface", required_argument, 0, 'i'},
616 {"queue", required_argument, 0, 'q'},
617 {"poll", no_argument, 0, 'p'},
618 {"xdp-skb", no_argument, 0, 'S'},
619 {"xdp-native", no_argument, 0, 'N'},
620 {"interval", required_argument, 0, 'n'},
621 {"zero-copy", no_argument, 0, 'z'},
622 {"copy", no_argument, 0, 'c'},
623 {"frame-size", required_argument, 0, 'f'},
624 {"no-need-wakeup", no_argument, 0, 'm'},
625 {"unaligned", no_argument, 0, 'u'},
626 {"shared-umem", no_argument, 0, 'M'},
627 {"force", no_argument, 0, 'F'},
628 {"duration", required_argument, 0, 'd'},
629 {"batch-size", required_argument, 0, 'b'},
630 {"tx-pkt-count", required_argument, 0, 'C'},
631 {"tx-pkt-size", required_argument, 0, 's'},
632 {"tx-pkt-pattern", required_argument, 0, 'P'},
636 static void usage(const char *prog)
639 " Usage: %s [OPTIONS]\n"
641 " -r, --rxdrop Discard all incoming packets (default)\n"
642 " -t, --txonly Only send packets\n"
643 " -l, --l2fwd MAC swap L2 forwarding\n"
644 " -i, --interface=n Run on interface n\n"
645 " -q, --queue=n Use queue n (default 0)\n"
646 " -p, --poll Use poll syscall\n"
647 " -S, --xdp-skb=n Use XDP skb-mod\n"
648 " -N, --xdp-native=n Enforce XDP native mode\n"
649 " -n, --interval=n Specify statistics update interval (default 1 sec).\n"
650 " -z, --zero-copy Force zero-copy mode.\n"
651 " -c, --copy Force copy mode.\n"
652 " -m, --no-need-wakeup Turn off use of driver need wakeup flag.\n"
653 " -f, --frame-size=n Set the frame size (must be a power of two in aligned mode, default is %d).\n"
654 " -u, --unaligned Enable unaligned chunk placement\n"
655 " -M, --shared-umem Enable XDP_SHARED_UMEM\n"
656 " -F, --force Force loading the XDP prog\n"
657 " -d, --duration=n Duration in secs to run command.\n"
658 " Default: forever.\n"
659 " -b, --batch-size=n Batch size for sending or receiving\n"
660 " packets. Default: %d\n"
661 " -C, --tx-pkt-count=n Number of packets to send.\n"
662 " Default: Continuous packets.\n"
663 " -s, --tx-pkt-size=n Transmit packet size.\n"
664 " (Default: %d bytes)\n"
665 " Min size: %d, Max size %d.\n"
666 " -P, --tx-pkt-pattern=nPacket fill pattern. Default: 0x%x\n"
668 fprintf(stderr, str, prog, XSK_UMEM__DEFAULT_FRAME_SIZE,
669 opt_batch_size, MIN_PKT_SIZE, MIN_PKT_SIZE,
670 XSK_UMEM__DEFAULT_FRAME_SIZE, opt_pkt_fill_pattern);
675 static void parse_command_line(int argc, char **argv)
682 c = getopt_long(argc, argv, "Frtli:q:pSNn:czf:muMd:b:C:s:P:",
683 long_options, &option_index);
689 opt_bench = BENCH_RXDROP;
692 opt_bench = BENCH_TXONLY;
695 opt_bench = BENCH_L2FWD;
701 opt_queue = atoi(optarg);
707 opt_xdp_flags |= XDP_FLAGS_SKB_MODE;
708 opt_xdp_bind_flags |= XDP_COPY;
711 /* default, set below */
714 opt_interval = atoi(optarg);
717 opt_xdp_bind_flags |= XDP_ZEROCOPY;
720 opt_xdp_bind_flags |= XDP_COPY;
723 opt_umem_flags |= XDP_UMEM_UNALIGNED_CHUNK_FLAG;
724 opt_unaligned_chunks = 1;
725 opt_mmap_flags = MAP_HUGETLB;
728 opt_xdp_flags &= ~XDP_FLAGS_UPDATE_IF_NOEXIST;
731 opt_xsk_frame_size = atoi(optarg);
734 opt_need_wakeup = false;
735 opt_xdp_bind_flags &= ~XDP_USE_NEED_WAKEUP;
738 opt_num_xsks = MAX_SOCKS;
741 opt_duration = atoi(optarg);
742 opt_duration *= 1000000000;
745 opt_batch_size = atoi(optarg);
748 opt_pkt_count = atoi(optarg);
751 opt_pkt_size = atoi(optarg);
752 if (opt_pkt_size > (XSK_UMEM__DEFAULT_FRAME_SIZE) ||
753 opt_pkt_size < MIN_PKT_SIZE) {
755 "ERROR: Invalid frame size %d\n",
757 usage(basename(argv[0]));
761 opt_pkt_fill_pattern = strtol(optarg, NULL, 16);
764 usage(basename(argv[0]));
768 if (!(opt_xdp_flags & XDP_FLAGS_SKB_MODE))
769 opt_xdp_flags |= XDP_FLAGS_DRV_MODE;
771 opt_ifindex = if_nametoindex(opt_if);
773 fprintf(stderr, "ERROR: interface \"%s\" does not exist\n",
775 usage(basename(argv[0]));
778 if ((opt_xsk_frame_size & (opt_xsk_frame_size - 1)) &&
779 !opt_unaligned_chunks) {
780 fprintf(stderr, "--frame-size=%d is not a power of two\n",
782 usage(basename(argv[0]));
786 static void kick_tx(struct xsk_socket_info *xsk)
790 ret = sendto(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, 0);
791 if (ret >= 0 || errno == ENOBUFS || errno == EAGAIN ||
792 errno == EBUSY || errno == ENETDOWN)
794 exit_with_error(errno);
797 static inline void complete_tx_l2fwd(struct xsk_socket_info *xsk,
800 struct xsk_umem_info *umem = xsk->umem;
801 u32 idx_cq = 0, idx_fq = 0;
805 if (!xsk->outstanding_tx)
808 if (!opt_need_wakeup || xsk_ring_prod__needs_wakeup(&xsk->tx))
811 ndescs = (xsk->outstanding_tx > opt_batch_size) ? opt_batch_size :
814 /* re-add completed Tx buffers */
815 rcvd = xsk_ring_cons__peek(&umem->cq, ndescs, &idx_cq);
820 ret = xsk_ring_prod__reserve(&umem->fq, rcvd, &idx_fq);
821 while (ret != rcvd) {
823 exit_with_error(-ret);
824 if (xsk_ring_prod__needs_wakeup(&umem->fq))
825 ret = poll(fds, num_socks, opt_timeout);
826 ret = xsk_ring_prod__reserve(&umem->fq, rcvd, &idx_fq);
829 for (i = 0; i < rcvd; i++)
830 *xsk_ring_prod__fill_addr(&umem->fq, idx_fq++) =
831 *xsk_ring_cons__comp_addr(&umem->cq, idx_cq++);
833 xsk_ring_prod__submit(&xsk->umem->fq, rcvd);
834 xsk_ring_cons__release(&xsk->umem->cq, rcvd);
835 xsk->outstanding_tx -= rcvd;
836 xsk->tx_npkts += rcvd;
840 static inline void complete_tx_only(struct xsk_socket_info *xsk,
846 if (!xsk->outstanding_tx)
849 if (!opt_need_wakeup || xsk_ring_prod__needs_wakeup(&xsk->tx))
852 rcvd = xsk_ring_cons__peek(&xsk->umem->cq, batch_size, &idx);
854 xsk_ring_cons__release(&xsk->umem->cq, rcvd);
855 xsk->outstanding_tx -= rcvd;
856 xsk->tx_npkts += rcvd;
860 static void rx_drop(struct xsk_socket_info *xsk, struct pollfd *fds)
862 unsigned int rcvd, i;
863 u32 idx_rx = 0, idx_fq = 0;
866 rcvd = xsk_ring_cons__peek(&xsk->rx, opt_batch_size, &idx_rx);
868 if (xsk_ring_prod__needs_wakeup(&xsk->umem->fq))
869 ret = poll(fds, num_socks, opt_timeout);
873 ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
874 while (ret != rcvd) {
876 exit_with_error(-ret);
877 if (xsk_ring_prod__needs_wakeup(&xsk->umem->fq))
878 ret = poll(fds, num_socks, opt_timeout);
879 ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
882 for (i = 0; i < rcvd; i++) {
883 u64 addr = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx)->addr;
884 u32 len = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx++)->len;
885 u64 orig = xsk_umem__extract_addr(addr);
887 addr = xsk_umem__add_offset_to_addr(addr);
888 char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
890 hex_dump(pkt, len, addr);
891 *xsk_ring_prod__fill_addr(&xsk->umem->fq, idx_fq++) = orig;
894 xsk_ring_prod__submit(&xsk->umem->fq, rcvd);
895 xsk_ring_cons__release(&xsk->rx, rcvd);
896 xsk->rx_npkts += rcvd;
899 static void rx_drop_all(void)
901 struct pollfd fds[MAX_SOCKS] = {};
904 for (i = 0; i < num_socks; i++) {
905 fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
906 fds[i].events = POLLIN;
911 ret = poll(fds, num_socks, opt_timeout);
916 for (i = 0; i < num_socks; i++)
917 rx_drop(xsks[i], fds);
924 static void tx_only(struct xsk_socket_info *xsk, u32 frame_nb, int batch_size)
929 while (xsk_ring_prod__reserve(&xsk->tx, batch_size, &idx) <
931 complete_tx_only(xsk, batch_size);
934 for (i = 0; i < batch_size; i++) {
935 struct xdp_desc *tx_desc = xsk_ring_prod__tx_desc(&xsk->tx,
937 tx_desc->addr = (frame_nb + i) << XSK_UMEM__DEFAULT_FRAME_SHIFT;
938 tx_desc->len = PKT_SIZE;
941 xsk_ring_prod__submit(&xsk->tx, batch_size);
942 xsk->outstanding_tx += batch_size;
943 frame_nb += batch_size;
944 frame_nb %= NUM_FRAMES;
945 complete_tx_only(xsk, batch_size);
948 static inline int get_batch_size(int pkt_cnt)
951 return opt_batch_size;
953 if (pkt_cnt + opt_batch_size <= opt_pkt_count)
954 return opt_batch_size;
956 return opt_pkt_count - pkt_cnt;
959 static void complete_tx_only_all(void)
966 for (i = 0; i < num_socks; i++) {
967 if (xsks[i]->outstanding_tx) {
968 complete_tx_only(xsks[i], opt_batch_size);
969 pending = !!xsks[i]->outstanding_tx;
975 static void tx_only_all(void)
977 struct pollfd fds[MAX_SOCKS] = {};
978 u32 frame_nb[MAX_SOCKS] = {};
982 for (i = 0; i < num_socks; i++) {
983 fds[0].fd = xsk_socket__fd(xsks[i]->xsk);
984 fds[0].events = POLLOUT;
987 while ((opt_pkt_count && pkt_cnt < opt_pkt_count) || !opt_pkt_count) {
988 int batch_size = get_batch_size(pkt_cnt);
991 ret = poll(fds, num_socks, opt_timeout);
995 if (!(fds[0].revents & POLLOUT))
999 for (i = 0; i < num_socks; i++)
1000 tx_only(xsks[i], frame_nb[i], batch_size);
1002 pkt_cnt += batch_size;
1009 complete_tx_only_all();
1012 static void l2fwd(struct xsk_socket_info *xsk, struct pollfd *fds)
1014 unsigned int rcvd, i;
1015 u32 idx_rx = 0, idx_tx = 0;
1018 complete_tx_l2fwd(xsk, fds);
1020 rcvd = xsk_ring_cons__peek(&xsk->rx, opt_batch_size, &idx_rx);
1022 if (xsk_ring_prod__needs_wakeup(&xsk->umem->fq))
1023 ret = poll(fds, num_socks, opt_timeout);
1027 ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
1028 while (ret != rcvd) {
1030 exit_with_error(-ret);
1031 if (xsk_ring_prod__needs_wakeup(&xsk->tx))
1033 ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
1036 for (i = 0; i < rcvd; i++) {
1037 u64 addr = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx)->addr;
1038 u32 len = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx++)->len;
1041 addr = xsk_umem__add_offset_to_addr(addr);
1042 char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
1044 swap_mac_addresses(pkt);
1046 hex_dump(pkt, len, addr);
1047 xsk_ring_prod__tx_desc(&xsk->tx, idx_tx)->addr = orig;
1048 xsk_ring_prod__tx_desc(&xsk->tx, idx_tx++)->len = len;
1051 xsk_ring_prod__submit(&xsk->tx, rcvd);
1052 xsk_ring_cons__release(&xsk->rx, rcvd);
1054 xsk->rx_npkts += rcvd;
1055 xsk->outstanding_tx += rcvd;
1058 static void l2fwd_all(void)
1060 struct pollfd fds[MAX_SOCKS] = {};
1063 for (i = 0; i < num_socks; i++) {
1064 fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
1065 fds[i].events = POLLOUT | POLLIN;
1070 ret = poll(fds, num_socks, opt_timeout);
1075 for (i = 0; i < num_socks; i++)
1076 l2fwd(xsks[i], fds);
1083 static void load_xdp_program(char **argv, struct bpf_object **obj)
1085 struct bpf_prog_load_attr prog_load_attr = {
1086 .prog_type = BPF_PROG_TYPE_XDP,
1088 char xdp_filename[256];
1091 snprintf(xdp_filename, sizeof(xdp_filename), "%s_kern.o", argv[0]);
1092 prog_load_attr.file = xdp_filename;
1094 if (bpf_prog_load_xattr(&prog_load_attr, obj, &prog_fd))
1097 fprintf(stderr, "ERROR: no program found: %s\n",
1102 if (bpf_set_link_xdp_fd(opt_ifindex, prog_fd, opt_xdp_flags) < 0) {
1103 fprintf(stderr, "ERROR: link set xdp fd failed\n");
1108 static void enter_xsks_into_map(struct bpf_object *obj)
1110 struct bpf_map *map;
1113 map = bpf_object__find_map_by_name(obj, "xsks_map");
1114 xsks_map = bpf_map__fd(map);
1116 fprintf(stderr, "ERROR: no xsks map found: %s\n",
1117 strerror(xsks_map));
1121 for (i = 0; i < num_socks; i++) {
1122 int fd = xsk_socket__fd(xsks[i]->xsk);
1126 ret = bpf_map_update_elem(xsks_map, &key, &fd, 0);
1128 fprintf(stderr, "ERROR: bpf_map_update_elem %d\n", i);
1134 int main(int argc, char **argv)
1136 struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
1137 bool rx = false, tx = false;
1138 struct xsk_umem_info *umem;
1139 struct bpf_object *obj;
1144 parse_command_line(argc, argv);
1146 if (setrlimit(RLIMIT_MEMLOCK, &r)) {
1147 fprintf(stderr, "ERROR: setrlimit(RLIMIT_MEMLOCK) \"%s\"\n",
1152 if (opt_num_xsks > 1)
1153 load_xdp_program(argv, &obj);
1155 /* Reserve memory for the umem. Use hugepages if unaligned chunk mode */
1156 bufs = mmap(NULL, NUM_FRAMES * opt_xsk_frame_size,
1157 PROT_READ | PROT_WRITE,
1158 MAP_PRIVATE | MAP_ANONYMOUS | opt_mmap_flags, -1, 0);
1159 if (bufs == MAP_FAILED) {
1160 printf("ERROR: mmap failed\n");
1164 /* Create sockets... */
1165 umem = xsk_configure_umem(bufs, NUM_FRAMES * opt_xsk_frame_size);
1166 if (opt_bench == BENCH_RXDROP || opt_bench == BENCH_L2FWD) {
1168 xsk_populate_fill_ring(umem);
1170 if (opt_bench == BENCH_L2FWD || opt_bench == BENCH_TXONLY)
1172 for (i = 0; i < opt_num_xsks; i++)
1173 xsks[num_socks++] = xsk_configure_socket(umem, rx, tx);
1175 if (opt_bench == BENCH_TXONLY) {
1178 for (i = 0; i < NUM_FRAMES; i++)
1179 gen_eth_frame(umem, i * opt_xsk_frame_size);
1182 if (opt_num_xsks > 1 && opt_bench != BENCH_TXONLY)
1183 enter_xsks_into_map(obj);
1185 signal(SIGINT, int_exit);
1186 signal(SIGTERM, int_exit);
1187 signal(SIGABRT, int_exit);
1189 setlocale(LC_ALL, "");
1191 ret = pthread_create(&pt, NULL, poller, NULL);
1193 exit_with_error(ret);
1195 prev_time = get_nsecs();
1196 start_time = prev_time;
1198 if (opt_bench == BENCH_RXDROP)
1200 else if (opt_bench == BENCH_TXONLY)
1205 benchmark_done = true;
1207 pthread_join(pt, NULL);