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/limits.h>
15 #include <linux/udp.h>
16 #include <arpa/inet.h>
18 #include <net/ethernet.h>
27 #include <sys/capability.h>
29 #include <sys/resource.h>
30 #include <sys/socket.h>
31 #include <sys/types.h>
36 #include <bpf/libbpf.h>
53 #define NUM_FRAMES (4 * 1024)
54 #define MIN_PKT_SIZE 64
56 #define DEBUG_HEXDUMP 0
63 static unsigned long prev_time;
71 static enum benchmark_type opt_bench = BENCH_RXDROP;
72 static u32 opt_xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
73 static const char *opt_if = "";
74 static int opt_ifindex;
76 static unsigned long opt_duration;
77 static unsigned long start_time;
78 static bool benchmark_done;
79 static u32 opt_batch_size = 64;
80 static int opt_pkt_count;
81 static u16 opt_pkt_size = MIN_PKT_SIZE;
82 static u32 opt_pkt_fill_pattern = 0x12345678;
83 static bool opt_extra_stats;
84 static bool opt_quiet;
85 static bool opt_app_stats;
86 static const char *opt_irq_str = "";
88 static int irqs_at_init = -1;
90 static int opt_interval = 1;
91 static u32 opt_xdp_bind_flags = XDP_USE_NEED_WAKEUP;
92 static u32 opt_umem_flags;
93 static int opt_unaligned_chunks;
94 static int opt_mmap_flags;
95 static int opt_xsk_frame_size = XSK_UMEM__DEFAULT_FRAME_SIZE;
96 static int opt_timeout = 1000;
97 static bool opt_need_wakeup = true;
98 static u32 opt_num_xsks = 1;
100 static bool opt_busy_poll;
101 static bool opt_reduced_cap;
103 struct xsk_ring_stats {
104 unsigned long rx_npkts;
105 unsigned long tx_npkts;
106 unsigned long rx_dropped_npkts;
107 unsigned long rx_invalid_npkts;
108 unsigned long tx_invalid_npkts;
109 unsigned long rx_full_npkts;
110 unsigned long rx_fill_empty_npkts;
111 unsigned long tx_empty_npkts;
112 unsigned long prev_rx_npkts;
113 unsigned long prev_tx_npkts;
114 unsigned long prev_rx_dropped_npkts;
115 unsigned long prev_rx_invalid_npkts;
116 unsigned long prev_tx_invalid_npkts;
117 unsigned long prev_rx_full_npkts;
118 unsigned long prev_rx_fill_empty_npkts;
119 unsigned long prev_tx_empty_npkts;
122 struct xsk_driver_stats {
124 unsigned long prev_intrs;
127 struct xsk_app_stats {
128 unsigned long rx_empty_polls;
129 unsigned long fill_fail_polls;
130 unsigned long copy_tx_sendtos;
131 unsigned long tx_wakeup_sendtos;
132 unsigned long opt_polls;
133 unsigned long prev_rx_empty_polls;
134 unsigned long prev_fill_fail_polls;
135 unsigned long prev_copy_tx_sendtos;
136 unsigned long prev_tx_wakeup_sendtos;
137 unsigned long prev_opt_polls;
140 struct xsk_umem_info {
141 struct xsk_ring_prod fq;
142 struct xsk_ring_cons cq;
143 struct xsk_umem *umem;
147 struct xsk_socket_info {
148 struct xsk_ring_cons rx;
149 struct xsk_ring_prod tx;
150 struct xsk_umem_info *umem;
151 struct xsk_socket *xsk;
152 struct xsk_ring_stats ring_stats;
153 struct xsk_app_stats app_stats;
154 struct xsk_driver_stats drv_stats;
158 static int num_socks;
159 struct xsk_socket_info *xsks[MAX_SOCKS];
162 static unsigned long get_nsecs(void)
166 clock_gettime(CLOCK_MONOTONIC, &ts);
167 return ts.tv_sec * 1000000000UL + ts.tv_nsec;
170 static void print_benchmark(bool running)
172 const char *bench_str = "INVALID";
174 if (opt_bench == BENCH_RXDROP)
175 bench_str = "rxdrop";
176 else if (opt_bench == BENCH_TXONLY)
177 bench_str = "txonly";
178 else if (opt_bench == BENCH_L2FWD)
181 printf("%s:%d %s ", opt_if, opt_queue, bench_str);
182 if (opt_xdp_flags & XDP_FLAGS_SKB_MODE)
184 else if (opt_xdp_flags & XDP_FLAGS_DRV_MODE)
193 printf("running...");
198 static int xsk_get_xdp_stats(int fd, struct xsk_socket_info *xsk)
200 struct xdp_statistics stats;
204 optlen = sizeof(stats);
205 err = getsockopt(fd, SOL_XDP, XDP_STATISTICS, &stats, &optlen);
209 if (optlen == sizeof(struct xdp_statistics)) {
210 xsk->ring_stats.rx_dropped_npkts = stats.rx_dropped;
211 xsk->ring_stats.rx_invalid_npkts = stats.rx_invalid_descs;
212 xsk->ring_stats.tx_invalid_npkts = stats.tx_invalid_descs;
213 xsk->ring_stats.rx_full_npkts = stats.rx_ring_full;
214 xsk->ring_stats.rx_fill_empty_npkts = stats.rx_fill_ring_empty_descs;
215 xsk->ring_stats.tx_empty_npkts = stats.tx_ring_empty_descs;
222 static void dump_app_stats(long dt)
226 for (i = 0; i < num_socks && xsks[i]; i++) {
227 char *fmt = "%-18s %'-14.0f %'-14lu\n";
228 double rx_empty_polls_ps, fill_fail_polls_ps, copy_tx_sendtos_ps,
229 tx_wakeup_sendtos_ps, opt_polls_ps;
231 rx_empty_polls_ps = (xsks[i]->app_stats.rx_empty_polls -
232 xsks[i]->app_stats.prev_rx_empty_polls) * 1000000000. / dt;
233 fill_fail_polls_ps = (xsks[i]->app_stats.fill_fail_polls -
234 xsks[i]->app_stats.prev_fill_fail_polls) * 1000000000. / dt;
235 copy_tx_sendtos_ps = (xsks[i]->app_stats.copy_tx_sendtos -
236 xsks[i]->app_stats.prev_copy_tx_sendtos) * 1000000000. / dt;
237 tx_wakeup_sendtos_ps = (xsks[i]->app_stats.tx_wakeup_sendtos -
238 xsks[i]->app_stats.prev_tx_wakeup_sendtos)
240 opt_polls_ps = (xsks[i]->app_stats.opt_polls -
241 xsks[i]->app_stats.prev_opt_polls) * 1000000000. / dt;
243 printf("\n%-18s %-14s %-14s\n", "", "calls/s", "count");
244 printf(fmt, "rx empty polls", rx_empty_polls_ps, xsks[i]->app_stats.rx_empty_polls);
245 printf(fmt, "fill fail polls", fill_fail_polls_ps,
246 xsks[i]->app_stats.fill_fail_polls);
247 printf(fmt, "copy tx sendtos", copy_tx_sendtos_ps,
248 xsks[i]->app_stats.copy_tx_sendtos);
249 printf(fmt, "tx wakeup sendtos", tx_wakeup_sendtos_ps,
250 xsks[i]->app_stats.tx_wakeup_sendtos);
251 printf(fmt, "opt polls", opt_polls_ps, xsks[i]->app_stats.opt_polls);
253 xsks[i]->app_stats.prev_rx_empty_polls = xsks[i]->app_stats.rx_empty_polls;
254 xsks[i]->app_stats.prev_fill_fail_polls = xsks[i]->app_stats.fill_fail_polls;
255 xsks[i]->app_stats.prev_copy_tx_sendtos = xsks[i]->app_stats.copy_tx_sendtos;
256 xsks[i]->app_stats.prev_tx_wakeup_sendtos = xsks[i]->app_stats.tx_wakeup_sendtos;
257 xsks[i]->app_stats.prev_opt_polls = xsks[i]->app_stats.opt_polls;
261 static bool get_interrupt_number(void)
267 f_int_proc = fopen("/proc/interrupts", "r");
268 if (f_int_proc == NULL) {
269 printf("Failed to open /proc/interrupts.\n");
273 while (!feof(f_int_proc) && !found) {
274 /* Make sure to read a full line at a time */
275 if (fgets(line, sizeof(line), f_int_proc) == NULL ||
276 line[strlen(line) - 1] != '\n') {
277 printf("Error reading from interrupts file\n");
281 /* Extract interrupt number from line */
282 if (strstr(line, opt_irq_str) != NULL) {
294 static int get_irqs(void)
296 char count_path[PATH_MAX];
297 int total_intrs = -1;
301 snprintf(count_path, sizeof(count_path),
302 "/sys/kernel/irq/%i/per_cpu_count", irq_no);
303 f_count_proc = fopen(count_path, "r");
304 if (f_count_proc == NULL) {
305 printf("Failed to open %s\n", count_path);
309 if (fgets(line, sizeof(line), f_count_proc) == NULL ||
310 line[strlen(line) - 1] != '\n') {
311 printf("Error reading from %s\n", count_path);
313 static const char com[2] = ",";
317 token = strtok(line, com);
318 while (token != NULL) {
319 /* sum up interrupts across all cores */
320 total_intrs += atoi(token);
321 token = strtok(NULL, com);
325 fclose(f_count_proc);
330 static void dump_driver_stats(long dt)
334 for (i = 0; i < num_socks && xsks[i]; i++) {
335 char *fmt = "%-18s %'-14.0f %'-14lu\n";
337 int n_ints = get_irqs();
340 printf("error getting intr info for intr %i\n", irq_no);
343 xsks[i]->drv_stats.intrs = n_ints - irqs_at_init;
345 intrs_ps = (xsks[i]->drv_stats.intrs - xsks[i]->drv_stats.prev_intrs) *
348 printf("\n%-18s %-14s %-14s\n", "", "intrs/s", "count");
349 printf(fmt, "irqs", intrs_ps, xsks[i]->drv_stats.intrs);
351 xsks[i]->drv_stats.prev_intrs = xsks[i]->drv_stats.intrs;
355 static void dump_stats(void)
357 unsigned long now = get_nsecs();
358 long dt = now - prev_time;
363 for (i = 0; i < num_socks && xsks[i]; i++) {
364 char *fmt = "%-18s %'-14.0f %'-14lu\n";
365 double rx_pps, tx_pps, dropped_pps, rx_invalid_pps, full_pps, fill_empty_pps,
366 tx_invalid_pps, tx_empty_pps;
368 rx_pps = (xsks[i]->ring_stats.rx_npkts - xsks[i]->ring_stats.prev_rx_npkts) *
370 tx_pps = (xsks[i]->ring_stats.tx_npkts - xsks[i]->ring_stats.prev_tx_npkts) *
373 printf("\n sock%d@", i);
374 print_benchmark(false);
377 printf("%-18s %-14s %-14s %-14.2f\n", "", "pps", "pkts",
379 printf(fmt, "rx", rx_pps, xsks[i]->ring_stats.rx_npkts);
380 printf(fmt, "tx", tx_pps, xsks[i]->ring_stats.tx_npkts);
382 xsks[i]->ring_stats.prev_rx_npkts = xsks[i]->ring_stats.rx_npkts;
383 xsks[i]->ring_stats.prev_tx_npkts = xsks[i]->ring_stats.tx_npkts;
385 if (opt_extra_stats) {
386 if (!xsk_get_xdp_stats(xsk_socket__fd(xsks[i]->xsk), xsks[i])) {
387 dropped_pps = (xsks[i]->ring_stats.rx_dropped_npkts -
388 xsks[i]->ring_stats.prev_rx_dropped_npkts) *
390 rx_invalid_pps = (xsks[i]->ring_stats.rx_invalid_npkts -
391 xsks[i]->ring_stats.prev_rx_invalid_npkts) *
393 tx_invalid_pps = (xsks[i]->ring_stats.tx_invalid_npkts -
394 xsks[i]->ring_stats.prev_tx_invalid_npkts) *
396 full_pps = (xsks[i]->ring_stats.rx_full_npkts -
397 xsks[i]->ring_stats.prev_rx_full_npkts) *
399 fill_empty_pps = (xsks[i]->ring_stats.rx_fill_empty_npkts -
400 xsks[i]->ring_stats.prev_rx_fill_empty_npkts) *
402 tx_empty_pps = (xsks[i]->ring_stats.tx_empty_npkts -
403 xsks[i]->ring_stats.prev_tx_empty_npkts) *
406 printf(fmt, "rx dropped", dropped_pps,
407 xsks[i]->ring_stats.rx_dropped_npkts);
408 printf(fmt, "rx invalid", rx_invalid_pps,
409 xsks[i]->ring_stats.rx_invalid_npkts);
410 printf(fmt, "tx invalid", tx_invalid_pps,
411 xsks[i]->ring_stats.tx_invalid_npkts);
412 printf(fmt, "rx queue full", full_pps,
413 xsks[i]->ring_stats.rx_full_npkts);
414 printf(fmt, "fill ring empty", fill_empty_pps,
415 xsks[i]->ring_stats.rx_fill_empty_npkts);
416 printf(fmt, "tx ring empty", tx_empty_pps,
417 xsks[i]->ring_stats.tx_empty_npkts);
419 xsks[i]->ring_stats.prev_rx_dropped_npkts =
420 xsks[i]->ring_stats.rx_dropped_npkts;
421 xsks[i]->ring_stats.prev_rx_invalid_npkts =
422 xsks[i]->ring_stats.rx_invalid_npkts;
423 xsks[i]->ring_stats.prev_tx_invalid_npkts =
424 xsks[i]->ring_stats.tx_invalid_npkts;
425 xsks[i]->ring_stats.prev_rx_full_npkts =
426 xsks[i]->ring_stats.rx_full_npkts;
427 xsks[i]->ring_stats.prev_rx_fill_empty_npkts =
428 xsks[i]->ring_stats.rx_fill_empty_npkts;
429 xsks[i]->ring_stats.prev_tx_empty_npkts =
430 xsks[i]->ring_stats.tx_empty_npkts;
432 printf("%-15s\n", "Error retrieving extra stats");
440 dump_driver_stats(dt);
443 static bool is_benchmark_done(void)
445 if (opt_duration > 0) {
446 unsigned long dt = (get_nsecs() - start_time);
448 if (dt >= opt_duration)
449 benchmark_done = true;
451 return benchmark_done;
454 static void *poller(void *arg)
457 while (!is_benchmark_done()) {
465 static void remove_xdp_program(void)
467 u32 curr_prog_id = 0;
469 if (bpf_get_link_xdp_id(opt_ifindex, &curr_prog_id, opt_xdp_flags)) {
470 printf("bpf_get_link_xdp_id failed\n");
474 if (prog_id == curr_prog_id)
475 bpf_set_link_xdp_fd(opt_ifindex, -1, opt_xdp_flags);
476 else if (!curr_prog_id)
477 printf("couldn't find a prog id on a given interface\n");
479 printf("program on interface changed, not removing\n");
482 static void int_exit(int sig)
484 benchmark_done = true;
487 static void __exit_with_error(int error, const char *file, const char *func,
490 fprintf(stderr, "%s:%s:%i: errno: %d/\"%s\"\n", file, func,
491 line, error, strerror(error));
493 if (opt_num_xsks > 1)
494 remove_xdp_program();
498 #define exit_with_error(error) __exit_with_error(error, __FILE__, __func__, __LINE__)
500 static void xdpsock_cleanup(void)
502 struct xsk_umem *umem = xsks[0]->umem->umem;
503 int i, cmd = CLOSE_CONN;
506 for (i = 0; i < num_socks; i++)
507 xsk_socket__delete(xsks[i]->xsk);
508 (void)xsk_umem__delete(umem);
510 if (opt_reduced_cap) {
511 if (write(sock, &cmd, sizeof(int)) < 0)
512 exit_with_error(errno);
515 if (opt_num_xsks > 1)
516 remove_xdp_program();
519 static void swap_mac_addresses(void *data)
521 struct ether_header *eth = (struct ether_header *)data;
522 struct ether_addr *src_addr = (struct ether_addr *)ð->ether_shost;
523 struct ether_addr *dst_addr = (struct ether_addr *)ð->ether_dhost;
524 struct ether_addr tmp;
527 *src_addr = *dst_addr;
531 static void hex_dump(void *pkt, size_t length, u64 addr)
533 const unsigned char *address = (unsigned char *)pkt;
534 const unsigned char *line = address;
535 size_t line_size = 32;
543 sprintf(buf, "addr=%llu", addr);
544 printf("length = %zu\n", length);
545 printf("%s | ", buf);
546 while (length-- > 0) {
547 printf("%02X ", *address++);
548 if (!(++i % line_size) || (length == 0 && i % line_size)) {
550 while (i++ % line_size)
553 printf(" | "); /* right close */
554 while (line < address) {
556 printf("%c", (c < 33 || c == 255) ? 0x2E : c);
560 printf("%s | ", buf);
566 static void *memset32_htonl(void *dest, u32 val, u32 size)
568 u32 *ptr = (u32 *)dest;
573 for (i = 0; i < (size & (~0x3)); i += 4)
576 for (; i < size; i++)
577 ((char *)dest)[i] = ((char *)&val)[i & 3];
583 * This function code has been taken from
584 * Linux kernel lib/checksum.c
586 static inline unsigned short from32to16(unsigned int x)
588 /* add up 16-bit and 16-bit for 16+c bit */
589 x = (x & 0xffff) + (x >> 16);
591 x = (x & 0xffff) + (x >> 16);
596 * This function code has been taken from
597 * Linux kernel lib/checksum.c
599 static unsigned int do_csum(const unsigned char *buff, int len)
601 unsigned int result = 0;
606 odd = 1 & (unsigned long)buff;
608 #ifdef __LITTLE_ENDIAN
609 result += (*buff << 8);
617 if (2 & (unsigned long)buff) {
618 result += *(unsigned short *)buff;
623 const unsigned char *end = buff +
624 ((unsigned int)len & ~3);
625 unsigned int carry = 0;
628 unsigned int w = *(unsigned int *)buff;
633 carry = (w > result);
634 } while (buff < end);
636 result = (result & 0xffff) + (result >> 16);
639 result += *(unsigned short *)buff;
644 #ifdef __LITTLE_ENDIAN
647 result += (*buff << 8);
649 result = from32to16(result);
651 result = ((result >> 8) & 0xff) | ((result & 0xff) << 8);
656 __sum16 ip_fast_csum(const void *iph, unsigned int ihl);
659 * This is a version of ip_compute_csum() optimized for IP headers,
660 * which always checksum on 4 octet boundaries.
661 * This function code has been taken from
662 * Linux kernel lib/checksum.c
664 __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
666 return (__force __sum16)~do_csum(iph, ihl * 4);
670 * Fold a partial checksum
671 * This function code has been taken from
672 * Linux kernel include/asm-generic/checksum.h
674 static inline __sum16 csum_fold(__wsum csum)
676 u32 sum = (__force u32)csum;
678 sum = (sum & 0xffff) + (sum >> 16);
679 sum = (sum & 0xffff) + (sum >> 16);
680 return (__force __sum16)~sum;
684 * This function code has been taken from
685 * Linux kernel lib/checksum.c
687 static inline u32 from64to32(u64 x)
689 /* add up 32-bit and 32-bit for 32+c bit */
690 x = (x & 0xffffffff) + (x >> 32);
692 x = (x & 0xffffffff) + (x >> 32);
696 __wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
697 __u32 len, __u8 proto, __wsum sum);
700 * This function code has been taken from
701 * Linux kernel lib/checksum.c
703 __wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
704 __u32 len, __u8 proto, __wsum sum)
706 unsigned long long s = (__force u32)sum;
708 s += (__force u32)saddr;
709 s += (__force u32)daddr;
710 #ifdef __BIG_ENDIAN__
713 s += (proto + len) << 8;
715 return (__force __wsum)from64to32(s);
719 * This function has been taken from
720 * Linux kernel include/asm-generic/checksum.h
722 static inline __sum16
723 csum_tcpudp_magic(__be32 saddr, __be32 daddr, __u32 len,
724 __u8 proto, __wsum sum)
726 return csum_fold(csum_tcpudp_nofold(saddr, daddr, len, proto, sum));
729 static inline u16 udp_csum(u32 saddr, u32 daddr, u32 len,
730 u8 proto, u16 *udp_pkt)
735 /* udp hdr and data */
736 for (; cnt < len; cnt += 2)
737 csum += udp_pkt[cnt >> 1];
739 return csum_tcpudp_magic(saddr, daddr, len, proto, csum);
742 #define ETH_FCS_SIZE 4
744 #define PKT_HDR_SIZE (sizeof(struct ethhdr) + sizeof(struct iphdr) + \
745 sizeof(struct udphdr))
747 #define PKT_SIZE (opt_pkt_size - ETH_FCS_SIZE)
748 #define IP_PKT_SIZE (PKT_SIZE - sizeof(struct ethhdr))
749 #define UDP_PKT_SIZE (IP_PKT_SIZE - sizeof(struct iphdr))
750 #define UDP_PKT_DATA_SIZE (UDP_PKT_SIZE - sizeof(struct udphdr))
752 static u8 pkt_data[XSK_UMEM__DEFAULT_FRAME_SIZE];
754 static void gen_eth_hdr_data(void)
756 struct udphdr *udp_hdr = (struct udphdr *)(pkt_data +
757 sizeof(struct ethhdr) +
758 sizeof(struct iphdr));
759 struct iphdr *ip_hdr = (struct iphdr *)(pkt_data +
760 sizeof(struct ethhdr));
761 struct ethhdr *eth_hdr = (struct ethhdr *)pkt_data;
763 /* ethernet header */
764 memcpy(eth_hdr->h_dest, "\x3c\xfd\xfe\x9e\x7f\x71", ETH_ALEN);
765 memcpy(eth_hdr->h_source, "\xec\xb1\xd7\x98\x3a\xc0", ETH_ALEN);
766 eth_hdr->h_proto = htons(ETH_P_IP);
769 ip_hdr->version = IPVERSION;
770 ip_hdr->ihl = 0x5; /* 20 byte header */
772 ip_hdr->tot_len = htons(IP_PKT_SIZE);
774 ip_hdr->frag_off = 0;
775 ip_hdr->ttl = IPDEFTTL;
776 ip_hdr->protocol = IPPROTO_UDP;
777 ip_hdr->saddr = htonl(0x0a0a0a10);
778 ip_hdr->daddr = htonl(0x0a0a0a20);
780 /* IP header checksum */
782 ip_hdr->check = ip_fast_csum((const void *)ip_hdr, ip_hdr->ihl);
785 udp_hdr->source = htons(0x1000);
786 udp_hdr->dest = htons(0x1000);
787 udp_hdr->len = htons(UDP_PKT_SIZE);
790 memset32_htonl(pkt_data + PKT_HDR_SIZE, opt_pkt_fill_pattern,
793 /* UDP header checksum */
795 udp_hdr->check = udp_csum(ip_hdr->saddr, ip_hdr->daddr, UDP_PKT_SIZE,
796 IPPROTO_UDP, (u16 *)udp_hdr);
799 static void gen_eth_frame(struct xsk_umem_info *umem, u64 addr)
801 memcpy(xsk_umem__get_data(umem->buffer, addr), pkt_data,
805 static struct xsk_umem_info *xsk_configure_umem(void *buffer, u64 size)
807 struct xsk_umem_info *umem;
808 struct xsk_umem_config cfg = {
809 /* We recommend that you set the fill ring size >= HW RX ring size +
810 * AF_XDP RX ring size. Make sure you fill up the fill ring
811 * with buffers at regular intervals, and you will with this setting
812 * avoid allocation failures in the driver. These are usually quite
813 * expensive since drivers have not been written to assume that
814 * allocation failures are common. For regular sockets, kernel
815 * allocated memory is used that only runs out in OOM situations
816 * that should be rare.
818 .fill_size = XSK_RING_PROD__DEFAULT_NUM_DESCS * 2,
819 .comp_size = XSK_RING_CONS__DEFAULT_NUM_DESCS,
820 .frame_size = opt_xsk_frame_size,
821 .frame_headroom = XSK_UMEM__DEFAULT_FRAME_HEADROOM,
822 .flags = opt_umem_flags
826 umem = calloc(1, sizeof(*umem));
828 exit_with_error(errno);
830 ret = xsk_umem__create(&umem->umem, buffer, size, &umem->fq, &umem->cq,
833 exit_with_error(-ret);
835 umem->buffer = buffer;
839 static void xsk_populate_fill_ring(struct xsk_umem_info *umem)
844 ret = xsk_ring_prod__reserve(&umem->fq,
845 XSK_RING_PROD__DEFAULT_NUM_DESCS * 2, &idx);
846 if (ret != XSK_RING_PROD__DEFAULT_NUM_DESCS * 2)
847 exit_with_error(-ret);
848 for (i = 0; i < XSK_RING_PROD__DEFAULT_NUM_DESCS * 2; i++)
849 *xsk_ring_prod__fill_addr(&umem->fq, idx++) =
850 i * opt_xsk_frame_size;
851 xsk_ring_prod__submit(&umem->fq, XSK_RING_PROD__DEFAULT_NUM_DESCS * 2);
854 static struct xsk_socket_info *xsk_configure_socket(struct xsk_umem_info *umem,
857 struct xsk_socket_config cfg;
858 struct xsk_socket_info *xsk;
859 struct xsk_ring_cons *rxr;
860 struct xsk_ring_prod *txr;
863 xsk = calloc(1, sizeof(*xsk));
865 exit_with_error(errno);
868 cfg.rx_size = XSK_RING_CONS__DEFAULT_NUM_DESCS;
869 cfg.tx_size = XSK_RING_PROD__DEFAULT_NUM_DESCS;
870 if (opt_num_xsks > 1 || opt_reduced_cap)
871 cfg.libbpf_flags = XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD;
873 cfg.libbpf_flags = 0;
874 cfg.xdp_flags = opt_xdp_flags;
875 cfg.bind_flags = opt_xdp_bind_flags;
877 rxr = rx ? &xsk->rx : NULL;
878 txr = tx ? &xsk->tx : NULL;
879 ret = xsk_socket__create(&xsk->xsk, opt_if, opt_queue, umem->umem,
882 exit_with_error(-ret);
884 ret = bpf_get_link_xdp_id(opt_ifindex, &prog_id, opt_xdp_flags);
886 exit_with_error(-ret);
888 xsk->app_stats.rx_empty_polls = 0;
889 xsk->app_stats.fill_fail_polls = 0;
890 xsk->app_stats.copy_tx_sendtos = 0;
891 xsk->app_stats.tx_wakeup_sendtos = 0;
892 xsk->app_stats.opt_polls = 0;
893 xsk->app_stats.prev_rx_empty_polls = 0;
894 xsk->app_stats.prev_fill_fail_polls = 0;
895 xsk->app_stats.prev_copy_tx_sendtos = 0;
896 xsk->app_stats.prev_tx_wakeup_sendtos = 0;
897 xsk->app_stats.prev_opt_polls = 0;
902 static struct option long_options[] = {
903 {"rxdrop", no_argument, 0, 'r'},
904 {"txonly", no_argument, 0, 't'},
905 {"l2fwd", no_argument, 0, 'l'},
906 {"interface", required_argument, 0, 'i'},
907 {"queue", required_argument, 0, 'q'},
908 {"poll", no_argument, 0, 'p'},
909 {"xdp-skb", no_argument, 0, 'S'},
910 {"xdp-native", no_argument, 0, 'N'},
911 {"interval", required_argument, 0, 'n'},
912 {"zero-copy", no_argument, 0, 'z'},
913 {"copy", no_argument, 0, 'c'},
914 {"frame-size", required_argument, 0, 'f'},
915 {"no-need-wakeup", no_argument, 0, 'm'},
916 {"unaligned", no_argument, 0, 'u'},
917 {"shared-umem", no_argument, 0, 'M'},
918 {"force", no_argument, 0, 'F'},
919 {"duration", required_argument, 0, 'd'},
920 {"batch-size", required_argument, 0, 'b'},
921 {"tx-pkt-count", required_argument, 0, 'C'},
922 {"tx-pkt-size", required_argument, 0, 's'},
923 {"tx-pkt-pattern", required_argument, 0, 'P'},
924 {"extra-stats", no_argument, 0, 'x'},
925 {"quiet", no_argument, 0, 'Q'},
926 {"app-stats", no_argument, 0, 'a'},
927 {"irq-string", no_argument, 0, 'I'},
928 {"busy-poll", no_argument, 0, 'B'},
929 {"reduce-cap", no_argument, 0, 'R'},
933 static void usage(const char *prog)
936 " Usage: %s [OPTIONS]\n"
938 " -r, --rxdrop Discard all incoming packets (default)\n"
939 " -t, --txonly Only send packets\n"
940 " -l, --l2fwd MAC swap L2 forwarding\n"
941 " -i, --interface=n Run on interface n\n"
942 " -q, --queue=n Use queue n (default 0)\n"
943 " -p, --poll Use poll syscall\n"
944 " -S, --xdp-skb=n Use XDP skb-mod\n"
945 " -N, --xdp-native=n Enforce XDP native mode\n"
946 " -n, --interval=n Specify statistics update interval (default 1 sec).\n"
947 " -z, --zero-copy Force zero-copy mode.\n"
948 " -c, --copy Force copy mode.\n"
949 " -m, --no-need-wakeup Turn off use of driver need wakeup flag.\n"
950 " -f, --frame-size=n Set the frame size (must be a power of two in aligned mode, default is %d).\n"
951 " -u, --unaligned Enable unaligned chunk placement\n"
952 " -M, --shared-umem Enable XDP_SHARED_UMEM (cannot be used with -R)\n"
953 " -F, --force Force loading the XDP prog\n"
954 " -d, --duration=n Duration in secs to run command.\n"
955 " Default: forever.\n"
956 " -b, --batch-size=n Batch size for sending or receiving\n"
957 " packets. Default: %d\n"
958 " -C, --tx-pkt-count=n Number of packets to send.\n"
959 " Default: Continuous packets.\n"
960 " -s, --tx-pkt-size=n Transmit packet size.\n"
961 " (Default: %d bytes)\n"
962 " Min size: %d, Max size %d.\n"
963 " -P, --tx-pkt-pattern=nPacket fill pattern. Default: 0x%x\n"
964 " -x, --extra-stats Display extra statistics.\n"
965 " -Q, --quiet Do not display any stats.\n"
966 " -a, --app-stats Display application (syscall) statistics.\n"
967 " -I, --irq-string Display driver interrupt statistics for interface associated with irq-string.\n"
968 " -B, --busy-poll Busy poll.\n"
969 " -R, --reduce-cap Use reduced capabilities (cannot be used with -M)\n"
971 fprintf(stderr, str, prog, XSK_UMEM__DEFAULT_FRAME_SIZE,
972 opt_batch_size, MIN_PKT_SIZE, MIN_PKT_SIZE,
973 XSK_UMEM__DEFAULT_FRAME_SIZE, opt_pkt_fill_pattern);
978 static void parse_command_line(int argc, char **argv)
985 c = getopt_long(argc, argv, "Frtli:q:pSNn:czf:muMd:b:C:s:P:xQaI:BR",
986 long_options, &option_index);
992 opt_bench = BENCH_RXDROP;
995 opt_bench = BENCH_TXONLY;
998 opt_bench = BENCH_L2FWD;
1004 opt_queue = atoi(optarg);
1010 opt_xdp_flags |= XDP_FLAGS_SKB_MODE;
1011 opt_xdp_bind_flags |= XDP_COPY;
1014 /* default, set below */
1017 opt_interval = atoi(optarg);
1020 opt_xdp_bind_flags |= XDP_ZEROCOPY;
1023 opt_xdp_bind_flags |= XDP_COPY;
1026 opt_umem_flags |= XDP_UMEM_UNALIGNED_CHUNK_FLAG;
1027 opt_unaligned_chunks = 1;
1028 opt_mmap_flags = MAP_HUGETLB;
1031 opt_xdp_flags &= ~XDP_FLAGS_UPDATE_IF_NOEXIST;
1034 opt_xsk_frame_size = atoi(optarg);
1037 opt_need_wakeup = false;
1038 opt_xdp_bind_flags &= ~XDP_USE_NEED_WAKEUP;
1041 opt_num_xsks = MAX_SOCKS;
1044 opt_duration = atoi(optarg);
1045 opt_duration *= 1000000000;
1048 opt_batch_size = atoi(optarg);
1051 opt_pkt_count = atoi(optarg);
1054 opt_pkt_size = atoi(optarg);
1055 if (opt_pkt_size > (XSK_UMEM__DEFAULT_FRAME_SIZE) ||
1056 opt_pkt_size < MIN_PKT_SIZE) {
1058 "ERROR: Invalid frame size %d\n",
1060 usage(basename(argv[0]));
1064 opt_pkt_fill_pattern = strtol(optarg, NULL, 16);
1067 opt_extra_stats = 1;
1076 opt_irq_str = optarg;
1077 if (get_interrupt_number())
1078 irqs_at_init = get_irqs();
1079 if (irqs_at_init < 0) {
1080 fprintf(stderr, "ERROR: Failed to get irqs for %s\n", opt_irq_str);
1081 usage(basename(argv[0]));
1088 opt_reduced_cap = true;
1091 usage(basename(argv[0]));
1095 if (!(opt_xdp_flags & XDP_FLAGS_SKB_MODE))
1096 opt_xdp_flags |= XDP_FLAGS_DRV_MODE;
1098 opt_ifindex = if_nametoindex(opt_if);
1100 fprintf(stderr, "ERROR: interface \"%s\" does not exist\n",
1102 usage(basename(argv[0]));
1105 if ((opt_xsk_frame_size & (opt_xsk_frame_size - 1)) &&
1106 !opt_unaligned_chunks) {
1107 fprintf(stderr, "--frame-size=%d is not a power of two\n",
1108 opt_xsk_frame_size);
1109 usage(basename(argv[0]));
1112 if (opt_reduced_cap && opt_num_xsks > 1) {
1113 fprintf(stderr, "ERROR: -M and -R cannot be used together\n");
1114 usage(basename(argv[0]));
1118 static void kick_tx(struct xsk_socket_info *xsk)
1122 ret = sendto(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, 0);
1123 if (ret >= 0 || errno == ENOBUFS || errno == EAGAIN ||
1124 errno == EBUSY || errno == ENETDOWN)
1126 exit_with_error(errno);
1129 static inline void complete_tx_l2fwd(struct xsk_socket_info *xsk)
1131 struct xsk_umem_info *umem = xsk->umem;
1132 u32 idx_cq = 0, idx_fq = 0;
1136 if (!xsk->outstanding_tx)
1139 /* In copy mode, Tx is driven by a syscall so we need to use e.g. sendto() to
1140 * really send the packets. In zero-copy mode we do not have to do this, since Tx
1141 * is driven by the NAPI loop. So as an optimization, we do not have to call
1142 * sendto() all the time in zero-copy mode for l2fwd.
1144 if (opt_xdp_bind_flags & XDP_COPY) {
1145 xsk->app_stats.copy_tx_sendtos++;
1149 ndescs = (xsk->outstanding_tx > opt_batch_size) ? opt_batch_size :
1150 xsk->outstanding_tx;
1152 /* re-add completed Tx buffers */
1153 rcvd = xsk_ring_cons__peek(&umem->cq, ndescs, &idx_cq);
1158 ret = xsk_ring_prod__reserve(&umem->fq, rcvd, &idx_fq);
1159 while (ret != rcvd) {
1161 exit_with_error(-ret);
1162 if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&umem->fq)) {
1163 xsk->app_stats.fill_fail_polls++;
1164 recvfrom(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL,
1167 ret = xsk_ring_prod__reserve(&umem->fq, rcvd, &idx_fq);
1170 for (i = 0; i < rcvd; i++)
1171 *xsk_ring_prod__fill_addr(&umem->fq, idx_fq++) =
1172 *xsk_ring_cons__comp_addr(&umem->cq, idx_cq++);
1174 xsk_ring_prod__submit(&xsk->umem->fq, rcvd);
1175 xsk_ring_cons__release(&xsk->umem->cq, rcvd);
1176 xsk->outstanding_tx -= rcvd;
1180 static inline void complete_tx_only(struct xsk_socket_info *xsk,
1186 if (!xsk->outstanding_tx)
1189 if (!opt_need_wakeup || xsk_ring_prod__needs_wakeup(&xsk->tx)) {
1190 xsk->app_stats.tx_wakeup_sendtos++;
1194 rcvd = xsk_ring_cons__peek(&xsk->umem->cq, batch_size, &idx);
1196 xsk_ring_cons__release(&xsk->umem->cq, rcvd);
1197 xsk->outstanding_tx -= rcvd;
1201 static void rx_drop(struct xsk_socket_info *xsk)
1203 unsigned int rcvd, i;
1204 u32 idx_rx = 0, idx_fq = 0;
1207 rcvd = xsk_ring_cons__peek(&xsk->rx, opt_batch_size, &idx_rx);
1209 if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&xsk->umem->fq)) {
1210 xsk->app_stats.rx_empty_polls++;
1211 recvfrom(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, NULL);
1216 ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
1217 while (ret != rcvd) {
1219 exit_with_error(-ret);
1220 if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&xsk->umem->fq)) {
1221 xsk->app_stats.fill_fail_polls++;
1222 recvfrom(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, NULL);
1224 ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
1227 for (i = 0; i < rcvd; i++) {
1228 u64 addr = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx)->addr;
1229 u32 len = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx++)->len;
1230 u64 orig = xsk_umem__extract_addr(addr);
1232 addr = xsk_umem__add_offset_to_addr(addr);
1233 char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
1235 hex_dump(pkt, len, addr);
1236 *xsk_ring_prod__fill_addr(&xsk->umem->fq, idx_fq++) = orig;
1239 xsk_ring_prod__submit(&xsk->umem->fq, rcvd);
1240 xsk_ring_cons__release(&xsk->rx, rcvd);
1241 xsk->ring_stats.rx_npkts += rcvd;
1244 static void rx_drop_all(void)
1246 struct pollfd fds[MAX_SOCKS] = {};
1249 for (i = 0; i < num_socks; i++) {
1250 fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
1251 fds[i].events = POLLIN;
1256 for (i = 0; i < num_socks; i++)
1257 xsks[i]->app_stats.opt_polls++;
1258 ret = poll(fds, num_socks, opt_timeout);
1263 for (i = 0; i < num_socks; i++)
1271 static void tx_only(struct xsk_socket_info *xsk, u32 *frame_nb, int batch_size)
1276 while (xsk_ring_prod__reserve(&xsk->tx, batch_size, &idx) <
1278 complete_tx_only(xsk, batch_size);
1283 for (i = 0; i < batch_size; i++) {
1284 struct xdp_desc *tx_desc = xsk_ring_prod__tx_desc(&xsk->tx,
1286 tx_desc->addr = (*frame_nb + i) * opt_xsk_frame_size;
1287 tx_desc->len = PKT_SIZE;
1290 xsk_ring_prod__submit(&xsk->tx, batch_size);
1291 xsk->ring_stats.tx_npkts += batch_size;
1292 xsk->outstanding_tx += batch_size;
1293 *frame_nb += batch_size;
1294 *frame_nb %= NUM_FRAMES;
1295 complete_tx_only(xsk, batch_size);
1298 static inline int get_batch_size(int pkt_cnt)
1301 return opt_batch_size;
1303 if (pkt_cnt + opt_batch_size <= opt_pkt_count)
1304 return opt_batch_size;
1306 return opt_pkt_count - pkt_cnt;
1309 static void complete_tx_only_all(void)
1316 for (i = 0; i < num_socks; i++) {
1317 if (xsks[i]->outstanding_tx) {
1318 complete_tx_only(xsks[i], opt_batch_size);
1319 pending = !!xsks[i]->outstanding_tx;
1325 static void tx_only_all(void)
1327 struct pollfd fds[MAX_SOCKS] = {};
1328 u32 frame_nb[MAX_SOCKS] = {};
1332 for (i = 0; i < num_socks; i++) {
1333 fds[0].fd = xsk_socket__fd(xsks[i]->xsk);
1334 fds[0].events = POLLOUT;
1337 while ((opt_pkt_count && pkt_cnt < opt_pkt_count) || !opt_pkt_count) {
1338 int batch_size = get_batch_size(pkt_cnt);
1341 for (i = 0; i < num_socks; i++)
1342 xsks[i]->app_stats.opt_polls++;
1343 ret = poll(fds, num_socks, opt_timeout);
1347 if (!(fds[0].revents & POLLOUT))
1351 for (i = 0; i < num_socks; i++)
1352 tx_only(xsks[i], &frame_nb[i], batch_size);
1354 pkt_cnt += batch_size;
1361 complete_tx_only_all();
1364 static void l2fwd(struct xsk_socket_info *xsk)
1366 unsigned int rcvd, i;
1367 u32 idx_rx = 0, idx_tx = 0;
1370 complete_tx_l2fwd(xsk);
1372 rcvd = xsk_ring_cons__peek(&xsk->rx, opt_batch_size, &idx_rx);
1374 if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&xsk->umem->fq)) {
1375 xsk->app_stats.rx_empty_polls++;
1376 recvfrom(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, NULL);
1380 xsk->ring_stats.rx_npkts += rcvd;
1382 ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
1383 while (ret != rcvd) {
1385 exit_with_error(-ret);
1386 complete_tx_l2fwd(xsk);
1387 if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&xsk->tx)) {
1388 xsk->app_stats.tx_wakeup_sendtos++;
1391 ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
1394 for (i = 0; i < rcvd; i++) {
1395 u64 addr = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx)->addr;
1396 u32 len = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx++)->len;
1399 addr = xsk_umem__add_offset_to_addr(addr);
1400 char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
1402 swap_mac_addresses(pkt);
1404 hex_dump(pkt, len, addr);
1405 xsk_ring_prod__tx_desc(&xsk->tx, idx_tx)->addr = orig;
1406 xsk_ring_prod__tx_desc(&xsk->tx, idx_tx++)->len = len;
1409 xsk_ring_prod__submit(&xsk->tx, rcvd);
1410 xsk_ring_cons__release(&xsk->rx, rcvd);
1412 xsk->ring_stats.tx_npkts += rcvd;
1413 xsk->outstanding_tx += rcvd;
1416 static void l2fwd_all(void)
1418 struct pollfd fds[MAX_SOCKS] = {};
1423 for (i = 0; i < num_socks; i++) {
1424 fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
1425 fds[i].events = POLLOUT | POLLIN;
1426 xsks[i]->app_stats.opt_polls++;
1428 ret = poll(fds, num_socks, opt_timeout);
1433 for (i = 0; i < num_socks; i++)
1441 static void load_xdp_program(char **argv, struct bpf_object **obj)
1443 struct bpf_prog_load_attr prog_load_attr = {
1444 .prog_type = BPF_PROG_TYPE_XDP,
1446 char xdp_filename[256];
1449 snprintf(xdp_filename, sizeof(xdp_filename), "%s_kern.o", argv[0]);
1450 prog_load_attr.file = xdp_filename;
1452 if (bpf_prog_load_xattr(&prog_load_attr, obj, &prog_fd))
1455 fprintf(stderr, "ERROR: no program found: %s\n",
1460 if (bpf_set_link_xdp_fd(opt_ifindex, prog_fd, opt_xdp_flags) < 0) {
1461 fprintf(stderr, "ERROR: link set xdp fd failed\n");
1466 static void enter_xsks_into_map(struct bpf_object *obj)
1468 struct bpf_map *map;
1471 map = bpf_object__find_map_by_name(obj, "xsks_map");
1472 xsks_map = bpf_map__fd(map);
1474 fprintf(stderr, "ERROR: no xsks map found: %s\n",
1475 strerror(xsks_map));
1479 for (i = 0; i < num_socks; i++) {
1480 int fd = xsk_socket__fd(xsks[i]->xsk);
1484 ret = bpf_map_update_elem(xsks_map, &key, &fd, 0);
1486 fprintf(stderr, "ERROR: bpf_map_update_elem %d\n", i);
1492 static void apply_setsockopt(struct xsk_socket_info *xsk)
1500 if (setsockopt(xsk_socket__fd(xsk->xsk), SOL_SOCKET, SO_PREFER_BUSY_POLL,
1501 (void *)&sock_opt, sizeof(sock_opt)) < 0)
1502 exit_with_error(errno);
1505 if (setsockopt(xsk_socket__fd(xsk->xsk), SOL_SOCKET, SO_BUSY_POLL,
1506 (void *)&sock_opt, sizeof(sock_opt)) < 0)
1507 exit_with_error(errno);
1509 sock_opt = opt_batch_size;
1510 if (setsockopt(xsk_socket__fd(xsk->xsk), SOL_SOCKET, SO_BUSY_POLL_BUDGET,
1511 (void *)&sock_opt, sizeof(sock_opt)) < 0)
1512 exit_with_error(errno);
1515 static int recv_xsks_map_fd_from_ctrl_node(int sock, int *_fd)
1517 char cms[CMSG_SPACE(sizeof(int))];
1518 struct cmsghdr *cmsg;
1524 iov.iov_base = &value;
1525 iov.iov_len = sizeof(int);
1528 msg.msg_namelen = 0;
1532 msg.msg_control = (caddr_t)cms;
1533 msg.msg_controllen = sizeof(cms);
1535 len = recvmsg(sock, &msg, 0);
1538 fprintf(stderr, "Recvmsg failed length incorrect.\n");
1543 fprintf(stderr, "Recvmsg failed no data\n");
1547 cmsg = CMSG_FIRSTHDR(&msg);
1548 *_fd = *(int *)CMSG_DATA(cmsg);
1554 recv_xsks_map_fd(int *xsks_map_fd)
1556 struct sockaddr_un server;
1559 sock = socket(AF_UNIX, SOCK_STREAM, 0);
1561 fprintf(stderr, "Error opening socket stream: %s", strerror(errno));
1565 server.sun_family = AF_UNIX;
1566 strcpy(server.sun_path, SOCKET_NAME);
1568 if (connect(sock, (struct sockaddr *)&server, sizeof(struct sockaddr_un)) < 0) {
1570 fprintf(stderr, "Error connecting stream socket: %s", strerror(errno));
1574 err = recv_xsks_map_fd_from_ctrl_node(sock, xsks_map_fd);
1576 fprintf(stderr, "Error %d receiving fd\n", err);
1582 int main(int argc, char **argv)
1584 struct __user_cap_header_struct hdr = { _LINUX_CAPABILITY_VERSION_3, 0 };
1585 struct __user_cap_data_struct data[2] = { { 0 } };
1586 struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
1587 bool rx = false, tx = false;
1588 struct xsk_umem_info *umem;
1589 struct bpf_object *obj;
1590 int xsks_map_fd = 0;
1595 parse_command_line(argc, argv);
1597 if (opt_reduced_cap) {
1598 if (capget(&hdr, data) < 0)
1599 fprintf(stderr, "Error getting capabilities\n");
1601 data->effective &= CAP_TO_MASK(CAP_NET_RAW);
1602 data->permitted &= CAP_TO_MASK(CAP_NET_RAW);
1604 if (capset(&hdr, data) < 0)
1605 fprintf(stderr, "Setting capabilities failed\n");
1607 if (capget(&hdr, data) < 0) {
1608 fprintf(stderr, "Error getting capabilities\n");
1610 fprintf(stderr, "Capabilities EFF %x Caps INH %x Caps Per %x\n",
1611 data[0].effective, data[0].inheritable, data[0].permitted);
1612 fprintf(stderr, "Capabilities EFF %x Caps INH %x Caps Per %x\n",
1613 data[1].effective, data[1].inheritable, data[1].permitted);
1616 if (setrlimit(RLIMIT_MEMLOCK, &r)) {
1617 fprintf(stderr, "ERROR: setrlimit(RLIMIT_MEMLOCK) \"%s\"\n",
1622 if (opt_num_xsks > 1)
1623 load_xdp_program(argv, &obj);
1626 /* Reserve memory for the umem. Use hugepages if unaligned chunk mode */
1627 bufs = mmap(NULL, NUM_FRAMES * opt_xsk_frame_size,
1628 PROT_READ | PROT_WRITE,
1629 MAP_PRIVATE | MAP_ANONYMOUS | opt_mmap_flags, -1, 0);
1630 if (bufs == MAP_FAILED) {
1631 printf("ERROR: mmap failed\n");
1635 /* Create sockets... */
1636 umem = xsk_configure_umem(bufs, NUM_FRAMES * opt_xsk_frame_size);
1637 if (opt_bench == BENCH_RXDROP || opt_bench == BENCH_L2FWD) {
1639 xsk_populate_fill_ring(umem);
1641 if (opt_bench == BENCH_L2FWD || opt_bench == BENCH_TXONLY)
1643 for (i = 0; i < opt_num_xsks; i++)
1644 xsks[num_socks++] = xsk_configure_socket(umem, rx, tx);
1646 for (i = 0; i < opt_num_xsks; i++)
1647 apply_setsockopt(xsks[i]);
1649 if (opt_bench == BENCH_TXONLY) {
1652 for (i = 0; i < NUM_FRAMES; i++)
1653 gen_eth_frame(umem, i * opt_xsk_frame_size);
1656 if (opt_num_xsks > 1 && opt_bench != BENCH_TXONLY)
1657 enter_xsks_into_map(obj);
1659 if (opt_reduced_cap) {
1660 ret = recv_xsks_map_fd(&xsks_map_fd);
1662 fprintf(stderr, "Error %d receiving xsks_map_fd\n", ret);
1663 exit_with_error(ret);
1666 ret = xsk_socket__update_xskmap(xsks[0]->xsk, xsks_map_fd);
1668 fprintf(stderr, "Update of BPF map failed(%d)\n", ret);
1669 exit_with_error(ret);
1674 signal(SIGINT, int_exit);
1675 signal(SIGTERM, int_exit);
1676 signal(SIGABRT, int_exit);
1678 setlocale(LC_ALL, "");
1681 ret = pthread_create(&pt, NULL, poller, NULL);
1683 exit_with_error(ret);
1686 prev_time = get_nsecs();
1687 start_time = prev_time;
1689 if (opt_bench == BENCH_RXDROP)
1691 else if (opt_bench == BENCH_TXONLY)
1696 benchmark_done = true;
1699 pthread_join(pt, NULL);
1703 munmap(bufs, NUM_FRAMES * opt_xsk_frame_size);