Loading...
Note: File does not exist in v6.13.7.
1// SPDX-License-Identifier: GPL-2.0
2/* Copyright(c) 2017 - 2018 Intel Corporation. */
3
4#include <asm/barrier.h>
5#include <errno.h>
6#include <getopt.h>
7#include <libgen.h>
8#include <linux/bpf.h>
9#include <linux/compiler.h>
10#include <linux/if_link.h>
11#include <linux/if_xdp.h>
12#include <linux/if_ether.h>
13#include <linux/ip.h>
14#include <linux/udp.h>
15#include <arpa/inet.h>
16#include <locale.h>
17#include <net/ethernet.h>
18#include <net/if.h>
19#include <poll.h>
20#include <pthread.h>
21#include <signal.h>
22#include <stdbool.h>
23#include <stdio.h>
24#include <stdlib.h>
25#include <string.h>
26#include <sys/mman.h>
27#include <sys/resource.h>
28#include <sys/socket.h>
29#include <sys/types.h>
30#include <time.h>
31#include <unistd.h>
32
33#include <bpf/libbpf.h>
34#include <bpf/xsk.h>
35#include <bpf/bpf.h>
36#include "xdpsock.h"
37
38#ifndef SOL_XDP
39#define SOL_XDP 283
40#endif
41
42#ifndef AF_XDP
43#define AF_XDP 44
44#endif
45
46#ifndef PF_XDP
47#define PF_XDP AF_XDP
48#endif
49
50#define NUM_FRAMES (4 * 1024)
51#define MIN_PKT_SIZE 64
52
53#define DEBUG_HEXDUMP 0
54
55typedef __u64 u64;
56typedef __u32 u32;
57typedef __u16 u16;
58typedef __u8 u8;
59
60static unsigned long prev_time;
61
62enum benchmark_type {
63 BENCH_RXDROP = 0,
64 BENCH_TXONLY = 1,
65 BENCH_L2FWD = 2,
66};
67
68static enum benchmark_type opt_bench = BENCH_RXDROP;
69static u32 opt_xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
70static const char *opt_if = "";
71static int opt_ifindex;
72static int opt_queue;
73static unsigned long opt_duration;
74static unsigned long start_time;
75static bool benchmark_done;
76static u32 opt_batch_size = 64;
77static int opt_pkt_count;
78static u16 opt_pkt_size = MIN_PKT_SIZE;
79static u32 opt_pkt_fill_pattern = 0x12345678;
80static bool opt_extra_stats;
81static int opt_poll;
82static int opt_interval = 1;
83static u32 opt_xdp_bind_flags = XDP_USE_NEED_WAKEUP;
84static u32 opt_umem_flags;
85static int opt_unaligned_chunks;
86static int opt_mmap_flags;
87static int opt_xsk_frame_size = XSK_UMEM__DEFAULT_FRAME_SIZE;
88static int opt_timeout = 1000;
89static bool opt_need_wakeup = true;
90static u32 opt_num_xsks = 1;
91static u32 prog_id;
92
93struct xsk_umem_info {
94 struct xsk_ring_prod fq;
95 struct xsk_ring_cons cq;
96 struct xsk_umem *umem;
97 void *buffer;
98};
99
100struct xsk_socket_info {
101 struct xsk_ring_cons rx;
102 struct xsk_ring_prod tx;
103 struct xsk_umem_info *umem;
104 struct xsk_socket *xsk;
105 unsigned long rx_npkts;
106 unsigned long tx_npkts;
107 unsigned long rx_dropped_npkts;
108 unsigned long rx_invalid_npkts;
109 unsigned long tx_invalid_npkts;
110 unsigned long rx_full_npkts;
111 unsigned long rx_fill_empty_npkts;
112 unsigned long tx_empty_npkts;
113 unsigned long prev_rx_npkts;
114 unsigned long prev_tx_npkts;
115 unsigned long prev_rx_dropped_npkts;
116 unsigned long prev_rx_invalid_npkts;
117 unsigned long prev_tx_invalid_npkts;
118 unsigned long prev_rx_full_npkts;
119 unsigned long prev_rx_fill_empty_npkts;
120 unsigned long prev_tx_empty_npkts;
121 u32 outstanding_tx;
122};
123
124static int num_socks;
125struct xsk_socket_info *xsks[MAX_SOCKS];
126
127static unsigned long get_nsecs(void)
128{
129 struct timespec ts;
130
131 clock_gettime(CLOCK_MONOTONIC, &ts);
132 return ts.tv_sec * 1000000000UL + ts.tv_nsec;
133}
134
135static void print_benchmark(bool running)
136{
137 const char *bench_str = "INVALID";
138
139 if (opt_bench == BENCH_RXDROP)
140 bench_str = "rxdrop";
141 else if (opt_bench == BENCH_TXONLY)
142 bench_str = "txonly";
143 else if (opt_bench == BENCH_L2FWD)
144 bench_str = "l2fwd";
145
146 printf("%s:%d %s ", opt_if, opt_queue, bench_str);
147 if (opt_xdp_flags & XDP_FLAGS_SKB_MODE)
148 printf("xdp-skb ");
149 else if (opt_xdp_flags & XDP_FLAGS_DRV_MODE)
150 printf("xdp-drv ");
151 else
152 printf(" ");
153
154 if (opt_poll)
155 printf("poll() ");
156
157 if (running) {
158 printf("running...");
159 fflush(stdout);
160 }
161}
162
163static int xsk_get_xdp_stats(int fd, struct xsk_socket_info *xsk)
164{
165 struct xdp_statistics stats;
166 socklen_t optlen;
167 int err;
168
169 optlen = sizeof(stats);
170 err = getsockopt(fd, SOL_XDP, XDP_STATISTICS, &stats, &optlen);
171 if (err)
172 return err;
173
174 if (optlen == sizeof(struct xdp_statistics)) {
175 xsk->rx_dropped_npkts = stats.rx_dropped;
176 xsk->rx_invalid_npkts = stats.rx_invalid_descs;
177 xsk->tx_invalid_npkts = stats.tx_invalid_descs;
178 xsk->rx_full_npkts = stats.rx_ring_full;
179 xsk->rx_fill_empty_npkts = stats.rx_fill_ring_empty_descs;
180 xsk->tx_empty_npkts = stats.tx_ring_empty_descs;
181 return 0;
182 }
183
184 return -EINVAL;
185}
186
187static void dump_stats(void)
188{
189 unsigned long now = get_nsecs();
190 long dt = now - prev_time;
191 int i;
192
193 prev_time = now;
194
195 for (i = 0; i < num_socks && xsks[i]; i++) {
196 char *fmt = "%-15s %'-11.0f %'-11lu\n";
197 double rx_pps, tx_pps, dropped_pps, rx_invalid_pps, full_pps, fill_empty_pps,
198 tx_invalid_pps, tx_empty_pps;
199
200 rx_pps = (xsks[i]->rx_npkts - xsks[i]->prev_rx_npkts) *
201 1000000000. / dt;
202 tx_pps = (xsks[i]->tx_npkts - xsks[i]->prev_tx_npkts) *
203 1000000000. / dt;
204
205 printf("\n sock%d@", i);
206 print_benchmark(false);
207 printf("\n");
208
209 printf("%-15s %-11s %-11s %-11.2f\n", "", "pps", "pkts",
210 dt / 1000000000.);
211 printf(fmt, "rx", rx_pps, xsks[i]->rx_npkts);
212 printf(fmt, "tx", tx_pps, xsks[i]->tx_npkts);
213
214 xsks[i]->prev_rx_npkts = xsks[i]->rx_npkts;
215 xsks[i]->prev_tx_npkts = xsks[i]->tx_npkts;
216
217 if (opt_extra_stats) {
218 if (!xsk_get_xdp_stats(xsk_socket__fd(xsks[i]->xsk), xsks[i])) {
219 dropped_pps = (xsks[i]->rx_dropped_npkts -
220 xsks[i]->prev_rx_dropped_npkts) * 1000000000. / dt;
221 rx_invalid_pps = (xsks[i]->rx_invalid_npkts -
222 xsks[i]->prev_rx_invalid_npkts) * 1000000000. / dt;
223 tx_invalid_pps = (xsks[i]->tx_invalid_npkts -
224 xsks[i]->prev_tx_invalid_npkts) * 1000000000. / dt;
225 full_pps = (xsks[i]->rx_full_npkts -
226 xsks[i]->prev_rx_full_npkts) * 1000000000. / dt;
227 fill_empty_pps = (xsks[i]->rx_fill_empty_npkts -
228 xsks[i]->prev_rx_fill_empty_npkts)
229 * 1000000000. / dt;
230 tx_empty_pps = (xsks[i]->tx_empty_npkts -
231 xsks[i]->prev_tx_empty_npkts) * 1000000000. / dt;
232
233 printf(fmt, "rx dropped", dropped_pps,
234 xsks[i]->rx_dropped_npkts);
235 printf(fmt, "rx invalid", rx_invalid_pps,
236 xsks[i]->rx_invalid_npkts);
237 printf(fmt, "tx invalid", tx_invalid_pps,
238 xsks[i]->tx_invalid_npkts);
239 printf(fmt, "rx queue full", full_pps,
240 xsks[i]->rx_full_npkts);
241 printf(fmt, "fill ring empty", fill_empty_pps,
242 xsks[i]->rx_fill_empty_npkts);
243 printf(fmt, "tx ring empty", tx_empty_pps,
244 xsks[i]->tx_empty_npkts);
245
246 xsks[i]->prev_rx_dropped_npkts = xsks[i]->rx_dropped_npkts;
247 xsks[i]->prev_rx_invalid_npkts = xsks[i]->rx_invalid_npkts;
248 xsks[i]->prev_tx_invalid_npkts = xsks[i]->tx_invalid_npkts;
249 xsks[i]->prev_rx_full_npkts = xsks[i]->rx_full_npkts;
250 xsks[i]->prev_rx_fill_empty_npkts = xsks[i]->rx_fill_empty_npkts;
251 xsks[i]->prev_tx_empty_npkts = xsks[i]->tx_empty_npkts;
252 } else {
253 printf("%-15s\n", "Error retrieving extra stats");
254 }
255 }
256 }
257}
258
259static bool is_benchmark_done(void)
260{
261 if (opt_duration > 0) {
262 unsigned long dt = (get_nsecs() - start_time);
263
264 if (dt >= opt_duration)
265 benchmark_done = true;
266 }
267 return benchmark_done;
268}
269
270static void *poller(void *arg)
271{
272 (void)arg;
273 while (!is_benchmark_done()) {
274 sleep(opt_interval);
275 dump_stats();
276 }
277
278 return NULL;
279}
280
281static void remove_xdp_program(void)
282{
283 u32 curr_prog_id = 0;
284
285 if (bpf_get_link_xdp_id(opt_ifindex, &curr_prog_id, opt_xdp_flags)) {
286 printf("bpf_get_link_xdp_id failed\n");
287 exit(EXIT_FAILURE);
288 }
289 if (prog_id == curr_prog_id)
290 bpf_set_link_xdp_fd(opt_ifindex, -1, opt_xdp_flags);
291 else if (!curr_prog_id)
292 printf("couldn't find a prog id on a given interface\n");
293 else
294 printf("program on interface changed, not removing\n");
295}
296
297static void int_exit(int sig)
298{
299 benchmark_done = true;
300}
301
302static void xdpsock_cleanup(void)
303{
304 struct xsk_umem *umem = xsks[0]->umem->umem;
305 int i;
306
307 dump_stats();
308 for (i = 0; i < num_socks; i++)
309 xsk_socket__delete(xsks[i]->xsk);
310 (void)xsk_umem__delete(umem);
311 remove_xdp_program();
312}
313
314static void __exit_with_error(int error, const char *file, const char *func,
315 int line)
316{
317 fprintf(stderr, "%s:%s:%i: errno: %d/\"%s\"\n", file, func,
318 line, error, strerror(error));
319 dump_stats();
320 remove_xdp_program();
321 exit(EXIT_FAILURE);
322}
323
324#define exit_with_error(error) __exit_with_error(error, __FILE__, __func__, \
325 __LINE__)
326static void swap_mac_addresses(void *data)
327{
328 struct ether_header *eth = (struct ether_header *)data;
329 struct ether_addr *src_addr = (struct ether_addr *)ð->ether_shost;
330 struct ether_addr *dst_addr = (struct ether_addr *)ð->ether_dhost;
331 struct ether_addr tmp;
332
333 tmp = *src_addr;
334 *src_addr = *dst_addr;
335 *dst_addr = tmp;
336}
337
338static void hex_dump(void *pkt, size_t length, u64 addr)
339{
340 const unsigned char *address = (unsigned char *)pkt;
341 const unsigned char *line = address;
342 size_t line_size = 32;
343 unsigned char c;
344 char buf[32];
345 int i = 0;
346
347 if (!DEBUG_HEXDUMP)
348 return;
349
350 sprintf(buf, "addr=%llu", addr);
351 printf("length = %zu\n", length);
352 printf("%s | ", buf);
353 while (length-- > 0) {
354 printf("%02X ", *address++);
355 if (!(++i % line_size) || (length == 0 && i % line_size)) {
356 if (length == 0) {
357 while (i++ % line_size)
358 printf("__ ");
359 }
360 printf(" | "); /* right close */
361 while (line < address) {
362 c = *line++;
363 printf("%c", (c < 33 || c == 255) ? 0x2E : c);
364 }
365 printf("\n");
366 if (length > 0)
367 printf("%s | ", buf);
368 }
369 }
370 printf("\n");
371}
372
373static void *memset32_htonl(void *dest, u32 val, u32 size)
374{
375 u32 *ptr = (u32 *)dest;
376 int i;
377
378 val = htonl(val);
379
380 for (i = 0; i < (size & (~0x3)); i += 4)
381 ptr[i >> 2] = val;
382
383 for (; i < size; i++)
384 ((char *)dest)[i] = ((char *)&val)[i & 3];
385
386 return dest;
387}
388
389/*
390 * This function code has been taken from
391 * Linux kernel lib/checksum.c
392 */
393static inline unsigned short from32to16(unsigned int x)
394{
395 /* add up 16-bit and 16-bit for 16+c bit */
396 x = (x & 0xffff) + (x >> 16);
397 /* add up carry.. */
398 x = (x & 0xffff) + (x >> 16);
399 return x;
400}
401
402/*
403 * This function code has been taken from
404 * Linux kernel lib/checksum.c
405 */
406static unsigned int do_csum(const unsigned char *buff, int len)
407{
408 unsigned int result = 0;
409 int odd;
410
411 if (len <= 0)
412 goto out;
413 odd = 1 & (unsigned long)buff;
414 if (odd) {
415#ifdef __LITTLE_ENDIAN
416 result += (*buff << 8);
417#else
418 result = *buff;
419#endif
420 len--;
421 buff++;
422 }
423 if (len >= 2) {
424 if (2 & (unsigned long)buff) {
425 result += *(unsigned short *)buff;
426 len -= 2;
427 buff += 2;
428 }
429 if (len >= 4) {
430 const unsigned char *end = buff +
431 ((unsigned int)len & ~3);
432 unsigned int carry = 0;
433
434 do {
435 unsigned int w = *(unsigned int *)buff;
436
437 buff += 4;
438 result += carry;
439 result += w;
440 carry = (w > result);
441 } while (buff < end);
442 result += carry;
443 result = (result & 0xffff) + (result >> 16);
444 }
445 if (len & 2) {
446 result += *(unsigned short *)buff;
447 buff += 2;
448 }
449 }
450 if (len & 1)
451#ifdef __LITTLE_ENDIAN
452 result += *buff;
453#else
454 result += (*buff << 8);
455#endif
456 result = from32to16(result);
457 if (odd)
458 result = ((result >> 8) & 0xff) | ((result & 0xff) << 8);
459out:
460 return result;
461}
462
463__sum16 ip_fast_csum(const void *iph, unsigned int ihl);
464
465/*
466 * This is a version of ip_compute_csum() optimized for IP headers,
467 * which always checksum on 4 octet boundaries.
468 * This function code has been taken from
469 * Linux kernel lib/checksum.c
470 */
471__sum16 ip_fast_csum(const void *iph, unsigned int ihl)
472{
473 return (__force __sum16)~do_csum(iph, ihl * 4);
474}
475
476/*
477 * Fold a partial checksum
478 * This function code has been taken from
479 * Linux kernel include/asm-generic/checksum.h
480 */
481static inline __sum16 csum_fold(__wsum csum)
482{
483 u32 sum = (__force u32)csum;
484
485 sum = (sum & 0xffff) + (sum >> 16);
486 sum = (sum & 0xffff) + (sum >> 16);
487 return (__force __sum16)~sum;
488}
489
490/*
491 * This function code has been taken from
492 * Linux kernel lib/checksum.c
493 */
494static inline u32 from64to32(u64 x)
495{
496 /* add up 32-bit and 32-bit for 32+c bit */
497 x = (x & 0xffffffff) + (x >> 32);
498 /* add up carry.. */
499 x = (x & 0xffffffff) + (x >> 32);
500 return (u32)x;
501}
502
503__wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
504 __u32 len, __u8 proto, __wsum sum);
505
506/*
507 * This function code has been taken from
508 * Linux kernel lib/checksum.c
509 */
510__wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
511 __u32 len, __u8 proto, __wsum sum)
512{
513 unsigned long long s = (__force u32)sum;
514
515 s += (__force u32)saddr;
516 s += (__force u32)daddr;
517#ifdef __BIG_ENDIAN__
518 s += proto + len;
519#else
520 s += (proto + len) << 8;
521#endif
522 return (__force __wsum)from64to32(s);
523}
524
525/*
526 * This function has been taken from
527 * Linux kernel include/asm-generic/checksum.h
528 */
529static inline __sum16
530csum_tcpudp_magic(__be32 saddr, __be32 daddr, __u32 len,
531 __u8 proto, __wsum sum)
532{
533 return csum_fold(csum_tcpudp_nofold(saddr, daddr, len, proto, sum));
534}
535
536static inline u16 udp_csum(u32 saddr, u32 daddr, u32 len,
537 u8 proto, u16 *udp_pkt)
538{
539 u32 csum = 0;
540 u32 cnt = 0;
541
542 /* udp hdr and data */
543 for (; cnt < len; cnt += 2)
544 csum += udp_pkt[cnt >> 1];
545
546 return csum_tcpudp_magic(saddr, daddr, len, proto, csum);
547}
548
549#define ETH_FCS_SIZE 4
550
551#define PKT_HDR_SIZE (sizeof(struct ethhdr) + sizeof(struct iphdr) + \
552 sizeof(struct udphdr))
553
554#define PKT_SIZE (opt_pkt_size - ETH_FCS_SIZE)
555#define IP_PKT_SIZE (PKT_SIZE - sizeof(struct ethhdr))
556#define UDP_PKT_SIZE (IP_PKT_SIZE - sizeof(struct iphdr))
557#define UDP_PKT_DATA_SIZE (UDP_PKT_SIZE - sizeof(struct udphdr))
558
559static u8 pkt_data[XSK_UMEM__DEFAULT_FRAME_SIZE];
560
561static void gen_eth_hdr_data(void)
562{
563 struct udphdr *udp_hdr = (struct udphdr *)(pkt_data +
564 sizeof(struct ethhdr) +
565 sizeof(struct iphdr));
566 struct iphdr *ip_hdr = (struct iphdr *)(pkt_data +
567 sizeof(struct ethhdr));
568 struct ethhdr *eth_hdr = (struct ethhdr *)pkt_data;
569
570 /* ethernet header */
571 memcpy(eth_hdr->h_dest, "\x3c\xfd\xfe\x9e\x7f\x71", ETH_ALEN);
572 memcpy(eth_hdr->h_source, "\xec\xb1\xd7\x98\x3a\xc0", ETH_ALEN);
573 eth_hdr->h_proto = htons(ETH_P_IP);
574
575 /* IP header */
576 ip_hdr->version = IPVERSION;
577 ip_hdr->ihl = 0x5; /* 20 byte header */
578 ip_hdr->tos = 0x0;
579 ip_hdr->tot_len = htons(IP_PKT_SIZE);
580 ip_hdr->id = 0;
581 ip_hdr->frag_off = 0;
582 ip_hdr->ttl = IPDEFTTL;
583 ip_hdr->protocol = IPPROTO_UDP;
584 ip_hdr->saddr = htonl(0x0a0a0a10);
585 ip_hdr->daddr = htonl(0x0a0a0a20);
586
587 /* IP header checksum */
588 ip_hdr->check = 0;
589 ip_hdr->check = ip_fast_csum((const void *)ip_hdr, ip_hdr->ihl);
590
591 /* UDP header */
592 udp_hdr->source = htons(0x1000);
593 udp_hdr->dest = htons(0x1000);
594 udp_hdr->len = htons(UDP_PKT_SIZE);
595
596 /* UDP data */
597 memset32_htonl(pkt_data + PKT_HDR_SIZE, opt_pkt_fill_pattern,
598 UDP_PKT_DATA_SIZE);
599
600 /* UDP header checksum */
601 udp_hdr->check = 0;
602 udp_hdr->check = udp_csum(ip_hdr->saddr, ip_hdr->daddr, UDP_PKT_SIZE,
603 IPPROTO_UDP, (u16 *)udp_hdr);
604}
605
606static void gen_eth_frame(struct xsk_umem_info *umem, u64 addr)
607{
608 memcpy(xsk_umem__get_data(umem->buffer, addr), pkt_data,
609 PKT_SIZE);
610}
611
612static struct xsk_umem_info *xsk_configure_umem(void *buffer, u64 size)
613{
614 struct xsk_umem_info *umem;
615 struct xsk_umem_config cfg = {
616 .fill_size = XSK_RING_PROD__DEFAULT_NUM_DESCS,
617 .comp_size = XSK_RING_CONS__DEFAULT_NUM_DESCS,
618 .frame_size = opt_xsk_frame_size,
619 .frame_headroom = XSK_UMEM__DEFAULT_FRAME_HEADROOM,
620 .flags = opt_umem_flags
621 };
622 int ret;
623
624 umem = calloc(1, sizeof(*umem));
625 if (!umem)
626 exit_with_error(errno);
627
628 ret = xsk_umem__create(&umem->umem, buffer, size, &umem->fq, &umem->cq,
629 &cfg);
630 if (ret)
631 exit_with_error(-ret);
632
633 umem->buffer = buffer;
634 return umem;
635}
636
637static void xsk_populate_fill_ring(struct xsk_umem_info *umem)
638{
639 int ret, i;
640 u32 idx;
641
642 ret = xsk_ring_prod__reserve(&umem->fq,
643 XSK_RING_PROD__DEFAULT_NUM_DESCS, &idx);
644 if (ret != XSK_RING_PROD__DEFAULT_NUM_DESCS)
645 exit_with_error(-ret);
646 for (i = 0; i < XSK_RING_PROD__DEFAULT_NUM_DESCS; i++)
647 *xsk_ring_prod__fill_addr(&umem->fq, idx++) =
648 i * opt_xsk_frame_size;
649 xsk_ring_prod__submit(&umem->fq, XSK_RING_PROD__DEFAULT_NUM_DESCS);
650}
651
652static struct xsk_socket_info *xsk_configure_socket(struct xsk_umem_info *umem,
653 bool rx, bool tx)
654{
655 struct xsk_socket_config cfg;
656 struct xsk_socket_info *xsk;
657 struct xsk_ring_cons *rxr;
658 struct xsk_ring_prod *txr;
659 int ret;
660
661 xsk = calloc(1, sizeof(*xsk));
662 if (!xsk)
663 exit_with_error(errno);
664
665 xsk->umem = umem;
666 cfg.rx_size = XSK_RING_CONS__DEFAULT_NUM_DESCS;
667 cfg.tx_size = XSK_RING_PROD__DEFAULT_NUM_DESCS;
668 if (opt_num_xsks > 1)
669 cfg.libbpf_flags = XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD;
670 else
671 cfg.libbpf_flags = 0;
672 cfg.xdp_flags = opt_xdp_flags;
673 cfg.bind_flags = opt_xdp_bind_flags;
674
675 rxr = rx ? &xsk->rx : NULL;
676 txr = tx ? &xsk->tx : NULL;
677 ret = xsk_socket__create(&xsk->xsk, opt_if, opt_queue, umem->umem,
678 rxr, txr, &cfg);
679 if (ret)
680 exit_with_error(-ret);
681
682 ret = bpf_get_link_xdp_id(opt_ifindex, &prog_id, opt_xdp_flags);
683 if (ret)
684 exit_with_error(-ret);
685
686 return xsk;
687}
688
689static struct option long_options[] = {
690 {"rxdrop", no_argument, 0, 'r'},
691 {"txonly", no_argument, 0, 't'},
692 {"l2fwd", no_argument, 0, 'l'},
693 {"interface", required_argument, 0, 'i'},
694 {"queue", required_argument, 0, 'q'},
695 {"poll", no_argument, 0, 'p'},
696 {"xdp-skb", no_argument, 0, 'S'},
697 {"xdp-native", no_argument, 0, 'N'},
698 {"interval", required_argument, 0, 'n'},
699 {"zero-copy", no_argument, 0, 'z'},
700 {"copy", no_argument, 0, 'c'},
701 {"frame-size", required_argument, 0, 'f'},
702 {"no-need-wakeup", no_argument, 0, 'm'},
703 {"unaligned", no_argument, 0, 'u'},
704 {"shared-umem", no_argument, 0, 'M'},
705 {"force", no_argument, 0, 'F'},
706 {"duration", required_argument, 0, 'd'},
707 {"batch-size", required_argument, 0, 'b'},
708 {"tx-pkt-count", required_argument, 0, 'C'},
709 {"tx-pkt-size", required_argument, 0, 's'},
710 {"tx-pkt-pattern", required_argument, 0, 'P'},
711 {"extra-stats", no_argument, 0, 'x'},
712 {0, 0, 0, 0}
713};
714
715static void usage(const char *prog)
716{
717 const char *str =
718 " Usage: %s [OPTIONS]\n"
719 " Options:\n"
720 " -r, --rxdrop Discard all incoming packets (default)\n"
721 " -t, --txonly Only send packets\n"
722 " -l, --l2fwd MAC swap L2 forwarding\n"
723 " -i, --interface=n Run on interface n\n"
724 " -q, --queue=n Use queue n (default 0)\n"
725 " -p, --poll Use poll syscall\n"
726 " -S, --xdp-skb=n Use XDP skb-mod\n"
727 " -N, --xdp-native=n Enforce XDP native mode\n"
728 " -n, --interval=n Specify statistics update interval (default 1 sec).\n"
729 " -z, --zero-copy Force zero-copy mode.\n"
730 " -c, --copy Force copy mode.\n"
731 " -m, --no-need-wakeup Turn off use of driver need wakeup flag.\n"
732 " -f, --frame-size=n Set the frame size (must be a power of two in aligned mode, default is %d).\n"
733 " -u, --unaligned Enable unaligned chunk placement\n"
734 " -M, --shared-umem Enable XDP_SHARED_UMEM\n"
735 " -F, --force Force loading the XDP prog\n"
736 " -d, --duration=n Duration in secs to run command.\n"
737 " Default: forever.\n"
738 " -b, --batch-size=n Batch size for sending or receiving\n"
739 " packets. Default: %d\n"
740 " -C, --tx-pkt-count=n Number of packets to send.\n"
741 " Default: Continuous packets.\n"
742 " -s, --tx-pkt-size=n Transmit packet size.\n"
743 " (Default: %d bytes)\n"
744 " Min size: %d, Max size %d.\n"
745 " -P, --tx-pkt-pattern=nPacket fill pattern. Default: 0x%x\n"
746 " -x, --extra-stats Display extra statistics.\n"
747 "\n";
748 fprintf(stderr, str, prog, XSK_UMEM__DEFAULT_FRAME_SIZE,
749 opt_batch_size, MIN_PKT_SIZE, MIN_PKT_SIZE,
750 XSK_UMEM__DEFAULT_FRAME_SIZE, opt_pkt_fill_pattern);
751
752 exit(EXIT_FAILURE);
753}
754
755static void parse_command_line(int argc, char **argv)
756{
757 int option_index, c;
758
759 opterr = 0;
760
761 for (;;) {
762 c = getopt_long(argc, argv, "Frtli:q:pSNn:czf:muMd:b:C:s:P:x",
763 long_options, &option_index);
764 if (c == -1)
765 break;
766
767 switch (c) {
768 case 'r':
769 opt_bench = BENCH_RXDROP;
770 break;
771 case 't':
772 opt_bench = BENCH_TXONLY;
773 break;
774 case 'l':
775 opt_bench = BENCH_L2FWD;
776 break;
777 case 'i':
778 opt_if = optarg;
779 break;
780 case 'q':
781 opt_queue = atoi(optarg);
782 break;
783 case 'p':
784 opt_poll = 1;
785 break;
786 case 'S':
787 opt_xdp_flags |= XDP_FLAGS_SKB_MODE;
788 opt_xdp_bind_flags |= XDP_COPY;
789 break;
790 case 'N':
791 /* default, set below */
792 break;
793 case 'n':
794 opt_interval = atoi(optarg);
795 break;
796 case 'z':
797 opt_xdp_bind_flags |= XDP_ZEROCOPY;
798 break;
799 case 'c':
800 opt_xdp_bind_flags |= XDP_COPY;
801 break;
802 case 'u':
803 opt_umem_flags |= XDP_UMEM_UNALIGNED_CHUNK_FLAG;
804 opt_unaligned_chunks = 1;
805 opt_mmap_flags = MAP_HUGETLB;
806 break;
807 case 'F':
808 opt_xdp_flags &= ~XDP_FLAGS_UPDATE_IF_NOEXIST;
809 break;
810 case 'f':
811 opt_xsk_frame_size = atoi(optarg);
812 break;
813 case 'm':
814 opt_need_wakeup = false;
815 opt_xdp_bind_flags &= ~XDP_USE_NEED_WAKEUP;
816 break;
817 case 'M':
818 opt_num_xsks = MAX_SOCKS;
819 break;
820 case 'd':
821 opt_duration = atoi(optarg);
822 opt_duration *= 1000000000;
823 break;
824 case 'b':
825 opt_batch_size = atoi(optarg);
826 break;
827 case 'C':
828 opt_pkt_count = atoi(optarg);
829 break;
830 case 's':
831 opt_pkt_size = atoi(optarg);
832 if (opt_pkt_size > (XSK_UMEM__DEFAULT_FRAME_SIZE) ||
833 opt_pkt_size < MIN_PKT_SIZE) {
834 fprintf(stderr,
835 "ERROR: Invalid frame size %d\n",
836 opt_pkt_size);
837 usage(basename(argv[0]));
838 }
839 break;
840 case 'P':
841 opt_pkt_fill_pattern = strtol(optarg, NULL, 16);
842 break;
843 case 'x':
844 opt_extra_stats = 1;
845 break;
846 default:
847 usage(basename(argv[0]));
848 }
849 }
850
851 if (!(opt_xdp_flags & XDP_FLAGS_SKB_MODE))
852 opt_xdp_flags |= XDP_FLAGS_DRV_MODE;
853
854 opt_ifindex = if_nametoindex(opt_if);
855 if (!opt_ifindex) {
856 fprintf(stderr, "ERROR: interface \"%s\" does not exist\n",
857 opt_if);
858 usage(basename(argv[0]));
859 }
860
861 if ((opt_xsk_frame_size & (opt_xsk_frame_size - 1)) &&
862 !opt_unaligned_chunks) {
863 fprintf(stderr, "--frame-size=%d is not a power of two\n",
864 opt_xsk_frame_size);
865 usage(basename(argv[0]));
866 }
867}
868
869static void kick_tx(struct xsk_socket_info *xsk)
870{
871 int ret;
872
873 ret = sendto(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, 0);
874 if (ret >= 0 || errno == ENOBUFS || errno == EAGAIN ||
875 errno == EBUSY || errno == ENETDOWN)
876 return;
877 exit_with_error(errno);
878}
879
880static inline void complete_tx_l2fwd(struct xsk_socket_info *xsk,
881 struct pollfd *fds)
882{
883 struct xsk_umem_info *umem = xsk->umem;
884 u32 idx_cq = 0, idx_fq = 0;
885 unsigned int rcvd;
886 size_t ndescs;
887
888 if (!xsk->outstanding_tx)
889 return;
890
891 if (!opt_need_wakeup || xsk_ring_prod__needs_wakeup(&xsk->tx))
892 kick_tx(xsk);
893
894 ndescs = (xsk->outstanding_tx > opt_batch_size) ? opt_batch_size :
895 xsk->outstanding_tx;
896
897 /* re-add completed Tx buffers */
898 rcvd = xsk_ring_cons__peek(&umem->cq, ndescs, &idx_cq);
899 if (rcvd > 0) {
900 unsigned int i;
901 int ret;
902
903 ret = xsk_ring_prod__reserve(&umem->fq, rcvd, &idx_fq);
904 while (ret != rcvd) {
905 if (ret < 0)
906 exit_with_error(-ret);
907 if (xsk_ring_prod__needs_wakeup(&umem->fq))
908 ret = poll(fds, num_socks, opt_timeout);
909 ret = xsk_ring_prod__reserve(&umem->fq, rcvd, &idx_fq);
910 }
911
912 for (i = 0; i < rcvd; i++)
913 *xsk_ring_prod__fill_addr(&umem->fq, idx_fq++) =
914 *xsk_ring_cons__comp_addr(&umem->cq, idx_cq++);
915
916 xsk_ring_prod__submit(&xsk->umem->fq, rcvd);
917 xsk_ring_cons__release(&xsk->umem->cq, rcvd);
918 xsk->outstanding_tx -= rcvd;
919 xsk->tx_npkts += rcvd;
920 }
921}
922
923static inline void complete_tx_only(struct xsk_socket_info *xsk,
924 int batch_size)
925{
926 unsigned int rcvd;
927 u32 idx;
928
929 if (!xsk->outstanding_tx)
930 return;
931
932 if (!opt_need_wakeup || xsk_ring_prod__needs_wakeup(&xsk->tx))
933 kick_tx(xsk);
934
935 rcvd = xsk_ring_cons__peek(&xsk->umem->cq, batch_size, &idx);
936 if (rcvd > 0) {
937 xsk_ring_cons__release(&xsk->umem->cq, rcvd);
938 xsk->outstanding_tx -= rcvd;
939 xsk->tx_npkts += rcvd;
940 }
941}
942
943static void rx_drop(struct xsk_socket_info *xsk, struct pollfd *fds)
944{
945 unsigned int rcvd, i;
946 u32 idx_rx = 0, idx_fq = 0;
947 int ret;
948
949 rcvd = xsk_ring_cons__peek(&xsk->rx, opt_batch_size, &idx_rx);
950 if (!rcvd) {
951 if (xsk_ring_prod__needs_wakeup(&xsk->umem->fq))
952 ret = poll(fds, num_socks, opt_timeout);
953 return;
954 }
955
956 ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
957 while (ret != rcvd) {
958 if (ret < 0)
959 exit_with_error(-ret);
960 if (xsk_ring_prod__needs_wakeup(&xsk->umem->fq))
961 ret = poll(fds, num_socks, opt_timeout);
962 ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
963 }
964
965 for (i = 0; i < rcvd; i++) {
966 u64 addr = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx)->addr;
967 u32 len = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx++)->len;
968 u64 orig = xsk_umem__extract_addr(addr);
969
970 addr = xsk_umem__add_offset_to_addr(addr);
971 char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
972
973 hex_dump(pkt, len, addr);
974 *xsk_ring_prod__fill_addr(&xsk->umem->fq, idx_fq++) = orig;
975 }
976
977 xsk_ring_prod__submit(&xsk->umem->fq, rcvd);
978 xsk_ring_cons__release(&xsk->rx, rcvd);
979 xsk->rx_npkts += rcvd;
980}
981
982static void rx_drop_all(void)
983{
984 struct pollfd fds[MAX_SOCKS] = {};
985 int i, ret;
986
987 for (i = 0; i < num_socks; i++) {
988 fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
989 fds[i].events = POLLIN;
990 }
991
992 for (;;) {
993 if (opt_poll) {
994 ret = poll(fds, num_socks, opt_timeout);
995 if (ret <= 0)
996 continue;
997 }
998
999 for (i = 0; i < num_socks; i++)
1000 rx_drop(xsks[i], fds);
1001
1002 if (benchmark_done)
1003 break;
1004 }
1005}
1006
1007static void tx_only(struct xsk_socket_info *xsk, u32 frame_nb, int batch_size)
1008{
1009 u32 idx;
1010 unsigned int i;
1011
1012 while (xsk_ring_prod__reserve(&xsk->tx, batch_size, &idx) <
1013 batch_size) {
1014 complete_tx_only(xsk, batch_size);
1015 }
1016
1017 for (i = 0; i < batch_size; i++) {
1018 struct xdp_desc *tx_desc = xsk_ring_prod__tx_desc(&xsk->tx,
1019 idx + i);
1020 tx_desc->addr = (frame_nb + i) << XSK_UMEM__DEFAULT_FRAME_SHIFT;
1021 tx_desc->len = PKT_SIZE;
1022 }
1023
1024 xsk_ring_prod__submit(&xsk->tx, batch_size);
1025 xsk->outstanding_tx += batch_size;
1026 frame_nb += batch_size;
1027 frame_nb %= NUM_FRAMES;
1028 complete_tx_only(xsk, batch_size);
1029}
1030
1031static inline int get_batch_size(int pkt_cnt)
1032{
1033 if (!opt_pkt_count)
1034 return opt_batch_size;
1035
1036 if (pkt_cnt + opt_batch_size <= opt_pkt_count)
1037 return opt_batch_size;
1038
1039 return opt_pkt_count - pkt_cnt;
1040}
1041
1042static void complete_tx_only_all(void)
1043{
1044 bool pending;
1045 int i;
1046
1047 do {
1048 pending = false;
1049 for (i = 0; i < num_socks; i++) {
1050 if (xsks[i]->outstanding_tx) {
1051 complete_tx_only(xsks[i], opt_batch_size);
1052 pending = !!xsks[i]->outstanding_tx;
1053 }
1054 }
1055 } while (pending);
1056}
1057
1058static void tx_only_all(void)
1059{
1060 struct pollfd fds[MAX_SOCKS] = {};
1061 u32 frame_nb[MAX_SOCKS] = {};
1062 int pkt_cnt = 0;
1063 int i, ret;
1064
1065 for (i = 0; i < num_socks; i++) {
1066 fds[0].fd = xsk_socket__fd(xsks[i]->xsk);
1067 fds[0].events = POLLOUT;
1068 }
1069
1070 while ((opt_pkt_count && pkt_cnt < opt_pkt_count) || !opt_pkt_count) {
1071 int batch_size = get_batch_size(pkt_cnt);
1072
1073 if (opt_poll) {
1074 ret = poll(fds, num_socks, opt_timeout);
1075 if (ret <= 0)
1076 continue;
1077
1078 if (!(fds[0].revents & POLLOUT))
1079 continue;
1080 }
1081
1082 for (i = 0; i < num_socks; i++)
1083 tx_only(xsks[i], frame_nb[i], batch_size);
1084
1085 pkt_cnt += batch_size;
1086
1087 if (benchmark_done)
1088 break;
1089 }
1090
1091 if (opt_pkt_count)
1092 complete_tx_only_all();
1093}
1094
1095static void l2fwd(struct xsk_socket_info *xsk, struct pollfd *fds)
1096{
1097 unsigned int rcvd, i;
1098 u32 idx_rx = 0, idx_tx = 0;
1099 int ret;
1100
1101 complete_tx_l2fwd(xsk, fds);
1102
1103 rcvd = xsk_ring_cons__peek(&xsk->rx, opt_batch_size, &idx_rx);
1104 if (!rcvd) {
1105 if (xsk_ring_prod__needs_wakeup(&xsk->umem->fq))
1106 ret = poll(fds, num_socks, opt_timeout);
1107 return;
1108 }
1109
1110 ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
1111 while (ret != rcvd) {
1112 if (ret < 0)
1113 exit_with_error(-ret);
1114 if (xsk_ring_prod__needs_wakeup(&xsk->tx))
1115 kick_tx(xsk);
1116 ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
1117 }
1118
1119 for (i = 0; i < rcvd; i++) {
1120 u64 addr = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx)->addr;
1121 u32 len = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx++)->len;
1122 u64 orig = addr;
1123
1124 addr = xsk_umem__add_offset_to_addr(addr);
1125 char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
1126
1127 swap_mac_addresses(pkt);
1128
1129 hex_dump(pkt, len, addr);
1130 xsk_ring_prod__tx_desc(&xsk->tx, idx_tx)->addr = orig;
1131 xsk_ring_prod__tx_desc(&xsk->tx, idx_tx++)->len = len;
1132 }
1133
1134 xsk_ring_prod__submit(&xsk->tx, rcvd);
1135 xsk_ring_cons__release(&xsk->rx, rcvd);
1136
1137 xsk->rx_npkts += rcvd;
1138 xsk->outstanding_tx += rcvd;
1139}
1140
1141static void l2fwd_all(void)
1142{
1143 struct pollfd fds[MAX_SOCKS] = {};
1144 int i, ret;
1145
1146 for (i = 0; i < num_socks; i++) {
1147 fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
1148 fds[i].events = POLLOUT | POLLIN;
1149 }
1150
1151 for (;;) {
1152 if (opt_poll) {
1153 ret = poll(fds, num_socks, opt_timeout);
1154 if (ret <= 0)
1155 continue;
1156 }
1157
1158 for (i = 0; i < num_socks; i++)
1159 l2fwd(xsks[i], fds);
1160
1161 if (benchmark_done)
1162 break;
1163 }
1164}
1165
1166static void load_xdp_program(char **argv, struct bpf_object **obj)
1167{
1168 struct bpf_prog_load_attr prog_load_attr = {
1169 .prog_type = BPF_PROG_TYPE_XDP,
1170 };
1171 char xdp_filename[256];
1172 int prog_fd;
1173
1174 snprintf(xdp_filename, sizeof(xdp_filename), "%s_kern.o", argv[0]);
1175 prog_load_attr.file = xdp_filename;
1176
1177 if (bpf_prog_load_xattr(&prog_load_attr, obj, &prog_fd))
1178 exit(EXIT_FAILURE);
1179 if (prog_fd < 0) {
1180 fprintf(stderr, "ERROR: no program found: %s\n",
1181 strerror(prog_fd));
1182 exit(EXIT_FAILURE);
1183 }
1184
1185 if (bpf_set_link_xdp_fd(opt_ifindex, prog_fd, opt_xdp_flags) < 0) {
1186 fprintf(stderr, "ERROR: link set xdp fd failed\n");
1187 exit(EXIT_FAILURE);
1188 }
1189}
1190
1191static void enter_xsks_into_map(struct bpf_object *obj)
1192{
1193 struct bpf_map *map;
1194 int i, xsks_map;
1195
1196 map = bpf_object__find_map_by_name(obj, "xsks_map");
1197 xsks_map = bpf_map__fd(map);
1198 if (xsks_map < 0) {
1199 fprintf(stderr, "ERROR: no xsks map found: %s\n",
1200 strerror(xsks_map));
1201 exit(EXIT_FAILURE);
1202 }
1203
1204 for (i = 0; i < num_socks; i++) {
1205 int fd = xsk_socket__fd(xsks[i]->xsk);
1206 int key, ret;
1207
1208 key = i;
1209 ret = bpf_map_update_elem(xsks_map, &key, &fd, 0);
1210 if (ret) {
1211 fprintf(stderr, "ERROR: bpf_map_update_elem %d\n", i);
1212 exit(EXIT_FAILURE);
1213 }
1214 }
1215}
1216
1217int main(int argc, char **argv)
1218{
1219 struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
1220 bool rx = false, tx = false;
1221 struct xsk_umem_info *umem;
1222 struct bpf_object *obj;
1223 pthread_t pt;
1224 int i, ret;
1225 void *bufs;
1226
1227 parse_command_line(argc, argv);
1228
1229 if (setrlimit(RLIMIT_MEMLOCK, &r)) {
1230 fprintf(stderr, "ERROR: setrlimit(RLIMIT_MEMLOCK) \"%s\"\n",
1231 strerror(errno));
1232 exit(EXIT_FAILURE);
1233 }
1234
1235 if (opt_num_xsks > 1)
1236 load_xdp_program(argv, &obj);
1237
1238 /* Reserve memory for the umem. Use hugepages if unaligned chunk mode */
1239 bufs = mmap(NULL, NUM_FRAMES * opt_xsk_frame_size,
1240 PROT_READ | PROT_WRITE,
1241 MAP_PRIVATE | MAP_ANONYMOUS | opt_mmap_flags, -1, 0);
1242 if (bufs == MAP_FAILED) {
1243 printf("ERROR: mmap failed\n");
1244 exit(EXIT_FAILURE);
1245 }
1246
1247 /* Create sockets... */
1248 umem = xsk_configure_umem(bufs, NUM_FRAMES * opt_xsk_frame_size);
1249 if (opt_bench == BENCH_RXDROP || opt_bench == BENCH_L2FWD) {
1250 rx = true;
1251 xsk_populate_fill_ring(umem);
1252 }
1253 if (opt_bench == BENCH_L2FWD || opt_bench == BENCH_TXONLY)
1254 tx = true;
1255 for (i = 0; i < opt_num_xsks; i++)
1256 xsks[num_socks++] = xsk_configure_socket(umem, rx, tx);
1257
1258 if (opt_bench == BENCH_TXONLY) {
1259 gen_eth_hdr_data();
1260
1261 for (i = 0; i < NUM_FRAMES; i++)
1262 gen_eth_frame(umem, i * opt_xsk_frame_size);
1263 }
1264
1265 if (opt_num_xsks > 1 && opt_bench != BENCH_TXONLY)
1266 enter_xsks_into_map(obj);
1267
1268 signal(SIGINT, int_exit);
1269 signal(SIGTERM, int_exit);
1270 signal(SIGABRT, int_exit);
1271
1272 setlocale(LC_ALL, "");
1273
1274 ret = pthread_create(&pt, NULL, poller, NULL);
1275 if (ret)
1276 exit_with_error(ret);
1277
1278 prev_time = get_nsecs();
1279 start_time = prev_time;
1280
1281 if (opt_bench == BENCH_RXDROP)
1282 rx_drop_all();
1283 else if (opt_bench == BENCH_TXONLY)
1284 tx_only_all();
1285 else
1286 l2fwd_all();
1287
1288 benchmark_done = true;
1289
1290 pthread_join(pt, NULL);
1291
1292 xdpsock_cleanup();
1293
1294 return 0;
1295}