• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2017 - 2018 Intel Corporation. */
3 
4 #include <errno.h>
5 #include <getopt.h>
6 #include <libgen.h>
7 #include <linux/bpf.h>
8 #include <linux/if_link.h>
9 #include <linux/if_xdp.h>
10 #include <linux/if_ether.h>
11 #include <linux/ip.h>
12 #include <linux/limits.h>
13 #include <linux/udp.h>
14 #include <arpa/inet.h>
15 #include <locale.h>
16 #include <net/ethernet.h>
17 #include <net/if.h>
18 #include <poll.h>
19 #include <pthread.h>
20 #include <signal.h>
21 #include <stdbool.h>
22 #include <stdio.h>
23 #include <stdlib.h>
24 #include <string.h>
25 #include <sys/capability.h>
26 #include <sys/mman.h>
27 #include <sys/resource.h>
28 #include <sys/socket.h>
29 #include <sys/types.h>
30 #include <sys/un.h>
31 #include <time.h>
32 #include <unistd.h>
33 
34 #include <bpf/libbpf.h>
35 #include <bpf/xsk.h>
36 #include <bpf/bpf.h>
37 #include "xdpsock.h"
38 
39 #ifndef SOL_XDP
40 #define SOL_XDP 283
41 #endif
42 
43 #ifndef AF_XDP
44 #define AF_XDP 44
45 #endif
46 
47 #ifndef PF_XDP
48 #define PF_XDP AF_XDP
49 #endif
50 
51 #define NUM_FRAMES (4 * 1024)
52 #define MIN_PKT_SIZE 64
53 
54 #define DEBUG_HEXDUMP 0
55 
56 typedef __u64 u64;
57 typedef __u32 u32;
58 typedef __u16 u16;
59 typedef __u8  u8;
60 
61 static unsigned long prev_time;
62 
63 enum benchmark_type {
64 	BENCH_RXDROP = 0,
65 	BENCH_TXONLY = 1,
66 	BENCH_L2FWD = 2,
67 };
68 
69 static enum benchmark_type opt_bench = BENCH_RXDROP;
70 static u32 opt_xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
71 static const char *opt_if = "";
72 static int opt_ifindex;
73 static int opt_queue;
74 static unsigned long opt_duration;
75 static unsigned long start_time;
76 static bool benchmark_done;
77 static u32 opt_batch_size = 64;
78 static int opt_pkt_count;
79 static u16 opt_pkt_size = MIN_PKT_SIZE;
80 static u32 opt_pkt_fill_pattern = 0x12345678;
81 static bool opt_extra_stats;
82 static bool opt_quiet;
83 static bool opt_app_stats;
84 static const char *opt_irq_str = "";
85 static u32 irq_no;
86 static int irqs_at_init = -1;
87 static int opt_poll;
88 static int opt_interval = 1;
89 static u32 opt_xdp_bind_flags = XDP_USE_NEED_WAKEUP;
90 static u32 opt_umem_flags;
91 static int opt_unaligned_chunks;
92 static int opt_mmap_flags;
93 static int opt_xsk_frame_size = XSK_UMEM__DEFAULT_FRAME_SIZE;
94 static int opt_timeout = 1000;
95 static bool opt_need_wakeup = true;
96 static u32 opt_num_xsks = 1;
97 static u32 prog_id;
98 static bool opt_busy_poll;
99 static bool opt_reduced_cap;
100 
101 struct xsk_ring_stats {
102 	unsigned long rx_npkts;
103 	unsigned long tx_npkts;
104 	unsigned long rx_dropped_npkts;
105 	unsigned long rx_invalid_npkts;
106 	unsigned long tx_invalid_npkts;
107 	unsigned long rx_full_npkts;
108 	unsigned long rx_fill_empty_npkts;
109 	unsigned long tx_empty_npkts;
110 	unsigned long prev_rx_npkts;
111 	unsigned long prev_tx_npkts;
112 	unsigned long prev_rx_dropped_npkts;
113 	unsigned long prev_rx_invalid_npkts;
114 	unsigned long prev_tx_invalid_npkts;
115 	unsigned long prev_rx_full_npkts;
116 	unsigned long prev_rx_fill_empty_npkts;
117 	unsigned long prev_tx_empty_npkts;
118 };
119 
120 struct xsk_driver_stats {
121 	unsigned long intrs;
122 	unsigned long prev_intrs;
123 };
124 
125 struct xsk_app_stats {
126 	unsigned long rx_empty_polls;
127 	unsigned long fill_fail_polls;
128 	unsigned long copy_tx_sendtos;
129 	unsigned long tx_wakeup_sendtos;
130 	unsigned long opt_polls;
131 	unsigned long prev_rx_empty_polls;
132 	unsigned long prev_fill_fail_polls;
133 	unsigned long prev_copy_tx_sendtos;
134 	unsigned long prev_tx_wakeup_sendtos;
135 	unsigned long prev_opt_polls;
136 };
137 
138 struct xsk_umem_info {
139 	struct xsk_ring_prod fq;
140 	struct xsk_ring_cons cq;
141 	struct xsk_umem *umem;
142 	void *buffer;
143 };
144 
145 struct xsk_socket_info {
146 	struct xsk_ring_cons rx;
147 	struct xsk_ring_prod tx;
148 	struct xsk_umem_info *umem;
149 	struct xsk_socket *xsk;
150 	struct xsk_ring_stats ring_stats;
151 	struct xsk_app_stats app_stats;
152 	struct xsk_driver_stats drv_stats;
153 	u32 outstanding_tx;
154 };
155 
156 static int num_socks;
157 struct xsk_socket_info *xsks[MAX_SOCKS];
158 int sock;
159 
get_nsecs(void)160 static unsigned long get_nsecs(void)
161 {
162 	struct timespec ts;
163 
164 	clock_gettime(CLOCK_MONOTONIC, &ts);
165 	return ts.tv_sec * 1000000000UL + ts.tv_nsec;
166 }
167 
print_benchmark(bool running)168 static void print_benchmark(bool running)
169 {
170 	const char *bench_str = "INVALID";
171 
172 	if (opt_bench == BENCH_RXDROP)
173 		bench_str = "rxdrop";
174 	else if (opt_bench == BENCH_TXONLY)
175 		bench_str = "txonly";
176 	else if (opt_bench == BENCH_L2FWD)
177 		bench_str = "l2fwd";
178 
179 	printf("%s:%d %s ", opt_if, opt_queue, bench_str);
180 	if (opt_xdp_flags & XDP_FLAGS_SKB_MODE)
181 		printf("xdp-skb ");
182 	else if (opt_xdp_flags & XDP_FLAGS_DRV_MODE)
183 		printf("xdp-drv ");
184 	else
185 		printf("	");
186 
187 	if (opt_poll)
188 		printf("poll() ");
189 
190 	if (running) {
191 		printf("running...");
192 		fflush(stdout);
193 	}
194 }
195 
xsk_get_xdp_stats(int fd,struct xsk_socket_info * xsk)196 static int xsk_get_xdp_stats(int fd, struct xsk_socket_info *xsk)
197 {
198 	struct xdp_statistics stats;
199 	socklen_t optlen;
200 	int err;
201 
202 	optlen = sizeof(stats);
203 	err = getsockopt(fd, SOL_XDP, XDP_STATISTICS, &stats, &optlen);
204 	if (err)
205 		return err;
206 
207 	if (optlen == sizeof(struct xdp_statistics)) {
208 		xsk->ring_stats.rx_dropped_npkts = stats.rx_dropped;
209 		xsk->ring_stats.rx_invalid_npkts = stats.rx_invalid_descs;
210 		xsk->ring_stats.tx_invalid_npkts = stats.tx_invalid_descs;
211 		xsk->ring_stats.rx_full_npkts = stats.rx_ring_full;
212 		xsk->ring_stats.rx_fill_empty_npkts = stats.rx_fill_ring_empty_descs;
213 		xsk->ring_stats.tx_empty_npkts = stats.tx_ring_empty_descs;
214 		return 0;
215 	}
216 
217 	return -EINVAL;
218 }
219 
dump_app_stats(long dt)220 static void dump_app_stats(long dt)
221 {
222 	int i;
223 
224 	for (i = 0; i < num_socks && xsks[i]; i++) {
225 		char *fmt = "%-18s %'-14.0f %'-14lu\n";
226 		double rx_empty_polls_ps, fill_fail_polls_ps, copy_tx_sendtos_ps,
227 				tx_wakeup_sendtos_ps, opt_polls_ps;
228 
229 		rx_empty_polls_ps = (xsks[i]->app_stats.rx_empty_polls -
230 					xsks[i]->app_stats.prev_rx_empty_polls) * 1000000000. / dt;
231 		fill_fail_polls_ps = (xsks[i]->app_stats.fill_fail_polls -
232 					xsks[i]->app_stats.prev_fill_fail_polls) * 1000000000. / dt;
233 		copy_tx_sendtos_ps = (xsks[i]->app_stats.copy_tx_sendtos -
234 					xsks[i]->app_stats.prev_copy_tx_sendtos) * 1000000000. / dt;
235 		tx_wakeup_sendtos_ps = (xsks[i]->app_stats.tx_wakeup_sendtos -
236 					xsks[i]->app_stats.prev_tx_wakeup_sendtos)
237 										* 1000000000. / dt;
238 		opt_polls_ps = (xsks[i]->app_stats.opt_polls -
239 					xsks[i]->app_stats.prev_opt_polls) * 1000000000. / dt;
240 
241 		printf("\n%-18s %-14s %-14s\n", "", "calls/s", "count");
242 		printf(fmt, "rx empty polls", rx_empty_polls_ps, xsks[i]->app_stats.rx_empty_polls);
243 		printf(fmt, "fill fail polls", fill_fail_polls_ps,
244 							xsks[i]->app_stats.fill_fail_polls);
245 		printf(fmt, "copy tx sendtos", copy_tx_sendtos_ps,
246 							xsks[i]->app_stats.copy_tx_sendtos);
247 		printf(fmt, "tx wakeup sendtos", tx_wakeup_sendtos_ps,
248 							xsks[i]->app_stats.tx_wakeup_sendtos);
249 		printf(fmt, "opt polls", opt_polls_ps, xsks[i]->app_stats.opt_polls);
250 
251 		xsks[i]->app_stats.prev_rx_empty_polls = xsks[i]->app_stats.rx_empty_polls;
252 		xsks[i]->app_stats.prev_fill_fail_polls = xsks[i]->app_stats.fill_fail_polls;
253 		xsks[i]->app_stats.prev_copy_tx_sendtos = xsks[i]->app_stats.copy_tx_sendtos;
254 		xsks[i]->app_stats.prev_tx_wakeup_sendtos = xsks[i]->app_stats.tx_wakeup_sendtos;
255 		xsks[i]->app_stats.prev_opt_polls = xsks[i]->app_stats.opt_polls;
256 	}
257 }
258 
get_interrupt_number(void)259 static bool get_interrupt_number(void)
260 {
261 	FILE *f_int_proc;
262 	char line[4096];
263 	bool found = false;
264 
265 	f_int_proc = fopen("/proc/interrupts", "r");
266 	if (f_int_proc == NULL) {
267 		printf("Failed to open /proc/interrupts.\n");
268 		return found;
269 	}
270 
271 	while (!feof(f_int_proc) && !found) {
272 		/* Make sure to read a full line at a time */
273 		if (fgets(line, sizeof(line), f_int_proc) == NULL ||
274 				line[strlen(line) - 1] != '\n') {
275 			printf("Error reading from interrupts file\n");
276 			break;
277 		}
278 
279 		/* Extract interrupt number from line */
280 		if (strstr(line, opt_irq_str) != NULL) {
281 			irq_no = atoi(line);
282 			found = true;
283 			break;
284 		}
285 	}
286 
287 	fclose(f_int_proc);
288 
289 	return found;
290 }
291 
get_irqs(void)292 static int get_irqs(void)
293 {
294 	char count_path[PATH_MAX];
295 	int total_intrs = -1;
296 	FILE *f_count_proc;
297 	char line[4096];
298 
299 	snprintf(count_path, sizeof(count_path),
300 		"/sys/kernel/irq/%i/per_cpu_count", irq_no);
301 	f_count_proc = fopen(count_path, "r");
302 	if (f_count_proc == NULL) {
303 		printf("Failed to open %s\n", count_path);
304 		return total_intrs;
305 	}
306 
307 	if (fgets(line, sizeof(line), f_count_proc) == NULL ||
308 			line[strlen(line) - 1] != '\n') {
309 		printf("Error reading from %s\n", count_path);
310 	} else {
311 		static const char com[2] = ",";
312 		char *token;
313 
314 		total_intrs = 0;
315 		token = strtok(line, com);
316 		while (token != NULL) {
317 			/* sum up interrupts across all cores */
318 			total_intrs += atoi(token);
319 			token = strtok(NULL, com);
320 		}
321 	}
322 
323 	fclose(f_count_proc);
324 
325 	return total_intrs;
326 }
327 
dump_driver_stats(long dt)328 static void dump_driver_stats(long dt)
329 {
330 	int i;
331 
332 	for (i = 0; i < num_socks && xsks[i]; i++) {
333 		char *fmt = "%-18s %'-14.0f %'-14lu\n";
334 		double intrs_ps;
335 		int n_ints = get_irqs();
336 
337 		if (n_ints < 0) {
338 			printf("error getting intr info for intr %i\n", irq_no);
339 			return;
340 		}
341 		xsks[i]->drv_stats.intrs = n_ints - irqs_at_init;
342 
343 		intrs_ps = (xsks[i]->drv_stats.intrs - xsks[i]->drv_stats.prev_intrs) *
344 			 1000000000. / dt;
345 
346 		printf("\n%-18s %-14s %-14s\n", "", "intrs/s", "count");
347 		printf(fmt, "irqs", intrs_ps, xsks[i]->drv_stats.intrs);
348 
349 		xsks[i]->drv_stats.prev_intrs = xsks[i]->drv_stats.intrs;
350 	}
351 }
352 
dump_stats(void)353 static void dump_stats(void)
354 {
355 	unsigned long now = get_nsecs();
356 	long dt = now - prev_time;
357 	int i;
358 
359 	prev_time = now;
360 
361 	for (i = 0; i < num_socks && xsks[i]; i++) {
362 		char *fmt = "%-18s %'-14.0f %'-14lu\n";
363 		double rx_pps, tx_pps, dropped_pps, rx_invalid_pps, full_pps, fill_empty_pps,
364 			tx_invalid_pps, tx_empty_pps;
365 
366 		rx_pps = (xsks[i]->ring_stats.rx_npkts - xsks[i]->ring_stats.prev_rx_npkts) *
367 			 1000000000. / dt;
368 		tx_pps = (xsks[i]->ring_stats.tx_npkts - xsks[i]->ring_stats.prev_tx_npkts) *
369 			 1000000000. / dt;
370 
371 		printf("\n sock%d@", i);
372 		print_benchmark(false);
373 		printf("\n");
374 
375 		printf("%-18s %-14s %-14s %-14.2f\n", "", "pps", "pkts",
376 		       dt / 1000000000.);
377 		printf(fmt, "rx", rx_pps, xsks[i]->ring_stats.rx_npkts);
378 		printf(fmt, "tx", tx_pps, xsks[i]->ring_stats.tx_npkts);
379 
380 		xsks[i]->ring_stats.prev_rx_npkts = xsks[i]->ring_stats.rx_npkts;
381 		xsks[i]->ring_stats.prev_tx_npkts = xsks[i]->ring_stats.tx_npkts;
382 
383 		if (opt_extra_stats) {
384 			if (!xsk_get_xdp_stats(xsk_socket__fd(xsks[i]->xsk), xsks[i])) {
385 				dropped_pps = (xsks[i]->ring_stats.rx_dropped_npkts -
386 						xsks[i]->ring_stats.prev_rx_dropped_npkts) *
387 							1000000000. / dt;
388 				rx_invalid_pps = (xsks[i]->ring_stats.rx_invalid_npkts -
389 						xsks[i]->ring_stats.prev_rx_invalid_npkts) *
390 							1000000000. / dt;
391 				tx_invalid_pps = (xsks[i]->ring_stats.tx_invalid_npkts -
392 						xsks[i]->ring_stats.prev_tx_invalid_npkts) *
393 							1000000000. / dt;
394 				full_pps = (xsks[i]->ring_stats.rx_full_npkts -
395 						xsks[i]->ring_stats.prev_rx_full_npkts) *
396 							1000000000. / dt;
397 				fill_empty_pps = (xsks[i]->ring_stats.rx_fill_empty_npkts -
398 						xsks[i]->ring_stats.prev_rx_fill_empty_npkts) *
399 							1000000000. / dt;
400 				tx_empty_pps = (xsks[i]->ring_stats.tx_empty_npkts -
401 						xsks[i]->ring_stats.prev_tx_empty_npkts) *
402 							1000000000. / dt;
403 
404 				printf(fmt, "rx dropped", dropped_pps,
405 				       xsks[i]->ring_stats.rx_dropped_npkts);
406 				printf(fmt, "rx invalid", rx_invalid_pps,
407 				       xsks[i]->ring_stats.rx_invalid_npkts);
408 				printf(fmt, "tx invalid", tx_invalid_pps,
409 				       xsks[i]->ring_stats.tx_invalid_npkts);
410 				printf(fmt, "rx queue full", full_pps,
411 				       xsks[i]->ring_stats.rx_full_npkts);
412 				printf(fmt, "fill ring empty", fill_empty_pps,
413 				       xsks[i]->ring_stats.rx_fill_empty_npkts);
414 				printf(fmt, "tx ring empty", tx_empty_pps,
415 				       xsks[i]->ring_stats.tx_empty_npkts);
416 
417 				xsks[i]->ring_stats.prev_rx_dropped_npkts =
418 					xsks[i]->ring_stats.rx_dropped_npkts;
419 				xsks[i]->ring_stats.prev_rx_invalid_npkts =
420 					xsks[i]->ring_stats.rx_invalid_npkts;
421 				xsks[i]->ring_stats.prev_tx_invalid_npkts =
422 					xsks[i]->ring_stats.tx_invalid_npkts;
423 				xsks[i]->ring_stats.prev_rx_full_npkts =
424 					xsks[i]->ring_stats.rx_full_npkts;
425 				xsks[i]->ring_stats.prev_rx_fill_empty_npkts =
426 					xsks[i]->ring_stats.rx_fill_empty_npkts;
427 				xsks[i]->ring_stats.prev_tx_empty_npkts =
428 					xsks[i]->ring_stats.tx_empty_npkts;
429 			} else {
430 				printf("%-15s\n", "Error retrieving extra stats");
431 			}
432 		}
433 	}
434 
435 	if (opt_app_stats)
436 		dump_app_stats(dt);
437 	if (irq_no)
438 		dump_driver_stats(dt);
439 }
440 
is_benchmark_done(void)441 static bool is_benchmark_done(void)
442 {
443 	if (opt_duration > 0) {
444 		unsigned long dt = (get_nsecs() - start_time);
445 
446 		if (dt >= opt_duration)
447 			benchmark_done = true;
448 	}
449 	return benchmark_done;
450 }
451 
poller(void * arg)452 static void *poller(void *arg)
453 {
454 	(void)arg;
455 	while (!is_benchmark_done()) {
456 		sleep(opt_interval);
457 		dump_stats();
458 	}
459 
460 	return NULL;
461 }
462 
remove_xdp_program(void)463 static void remove_xdp_program(void)
464 {
465 	u32 curr_prog_id = 0;
466 
467 	if (bpf_get_link_xdp_id(opt_ifindex, &curr_prog_id, opt_xdp_flags)) {
468 		printf("bpf_get_link_xdp_id failed\n");
469 		exit(EXIT_FAILURE);
470 	}
471 
472 	if (prog_id == curr_prog_id)
473 		bpf_set_link_xdp_fd(opt_ifindex, -1, opt_xdp_flags);
474 	else if (!curr_prog_id)
475 		printf("couldn't find a prog id on a given interface\n");
476 	else
477 		printf("program on interface changed, not removing\n");
478 }
479 
int_exit(int sig)480 static void int_exit(int sig)
481 {
482 	benchmark_done = true;
483 }
484 
__exit_with_error(int error,const char * file,const char * func,int line)485 static void __exit_with_error(int error, const char *file, const char *func,
486 			      int line)
487 {
488 	fprintf(stderr, "%s:%s:%i: errno: %d/\"%s\"\n", file, func,
489 		line, error, strerror(error));
490 
491 	if (opt_num_xsks > 1)
492 		remove_xdp_program();
493 	exit(EXIT_FAILURE);
494 }
495 
496 #define exit_with_error(error) __exit_with_error(error, __FILE__, __func__, __LINE__)
497 
xdpsock_cleanup(void)498 static void xdpsock_cleanup(void)
499 {
500 	struct xsk_umem *umem = xsks[0]->umem->umem;
501 	int i, cmd = CLOSE_CONN;
502 
503 	dump_stats();
504 	for (i = 0; i < num_socks; i++)
505 		xsk_socket__delete(xsks[i]->xsk);
506 	(void)xsk_umem__delete(umem);
507 
508 	if (opt_reduced_cap) {
509 		if (write(sock, &cmd, sizeof(int)) < 0)
510 			exit_with_error(errno);
511 	}
512 
513 	if (opt_num_xsks > 1)
514 		remove_xdp_program();
515 }
516 
swap_mac_addresses(void * data)517 static void swap_mac_addresses(void *data)
518 {
519 	struct ether_header *eth = (struct ether_header *)data;
520 	struct ether_addr *src_addr = (struct ether_addr *)&eth->ether_shost;
521 	struct ether_addr *dst_addr = (struct ether_addr *)&eth->ether_dhost;
522 	struct ether_addr tmp;
523 
524 	tmp = *src_addr;
525 	*src_addr = *dst_addr;
526 	*dst_addr = tmp;
527 }
528 
hex_dump(void * pkt,size_t length,u64 addr)529 static void hex_dump(void *pkt, size_t length, u64 addr)
530 {
531 	const unsigned char *address = (unsigned char *)pkt;
532 	const unsigned char *line = address;
533 	size_t line_size = 32;
534 	unsigned char c;
535 	char buf[32];
536 	int i = 0;
537 
538 	if (!DEBUG_HEXDUMP)
539 		return;
540 
541 	sprintf(buf, "addr=%llu", addr);
542 	printf("length = %zu\n", length);
543 	printf("%s | ", buf);
544 	while (length-- > 0) {
545 		printf("%02X ", *address++);
546 		if (!(++i % line_size) || (length == 0 && i % line_size)) {
547 			if (length == 0) {
548 				while (i++ % line_size)
549 					printf("__ ");
550 			}
551 			printf(" | ");	/* right close */
552 			while (line < address) {
553 				c = *line++;
554 				printf("%c", (c < 33 || c == 255) ? 0x2E : c);
555 			}
556 			printf("\n");
557 			if (length > 0)
558 				printf("%s | ", buf);
559 		}
560 	}
561 	printf("\n");
562 }
563 
memset32_htonl(void * dest,u32 val,u32 size)564 static void *memset32_htonl(void *dest, u32 val, u32 size)
565 {
566 	u32 *ptr = (u32 *)dest;
567 	int i;
568 
569 	val = htonl(val);
570 
571 	for (i = 0; i < (size & (~0x3)); i += 4)
572 		ptr[i >> 2] = val;
573 
574 	for (; i < size; i++)
575 		((char *)dest)[i] = ((char *)&val)[i & 3];
576 
577 	return dest;
578 }
579 
580 /*
581  * This function code has been taken from
582  * Linux kernel lib/checksum.c
583  */
from32to16(unsigned int x)584 static inline unsigned short from32to16(unsigned int x)
585 {
586 	/* add up 16-bit and 16-bit for 16+c bit */
587 	x = (x & 0xffff) + (x >> 16);
588 	/* add up carry.. */
589 	x = (x & 0xffff) + (x >> 16);
590 	return x;
591 }
592 
593 /*
594  * This function code has been taken from
595  * Linux kernel lib/checksum.c
596  */
do_csum(const unsigned char * buff,int len)597 static unsigned int do_csum(const unsigned char *buff, int len)
598 {
599 	unsigned int result = 0;
600 	int odd;
601 
602 	if (len <= 0)
603 		goto out;
604 	odd = 1 & (unsigned long)buff;
605 	if (odd) {
606 #ifdef __LITTLE_ENDIAN
607 		result += (*buff << 8);
608 #else
609 		result = *buff;
610 #endif
611 		len--;
612 		buff++;
613 	}
614 	if (len >= 2) {
615 		if (2 & (unsigned long)buff) {
616 			result += *(unsigned short *)buff;
617 			len -= 2;
618 			buff += 2;
619 		}
620 		if (len >= 4) {
621 			const unsigned char *end = buff +
622 						   ((unsigned int)len & ~3);
623 			unsigned int carry = 0;
624 
625 			do {
626 				unsigned int w = *(unsigned int *)buff;
627 
628 				buff += 4;
629 				result += carry;
630 				result += w;
631 				carry = (w > result);
632 			} while (buff < end);
633 			result += carry;
634 			result = (result & 0xffff) + (result >> 16);
635 		}
636 		if (len & 2) {
637 			result += *(unsigned short *)buff;
638 			buff += 2;
639 		}
640 	}
641 	if (len & 1)
642 #ifdef __LITTLE_ENDIAN
643 		result += *buff;
644 #else
645 		result += (*buff << 8);
646 #endif
647 	result = from32to16(result);
648 	if (odd)
649 		result = ((result >> 8) & 0xff) | ((result & 0xff) << 8);
650 out:
651 	return result;
652 }
653 
654 /*
655  *	This is a version of ip_compute_csum() optimized for IP headers,
656  *	which always checksum on 4 octet boundaries.
657  *	This function code has been taken from
658  *	Linux kernel lib/checksum.c
659  */
ip_fast_csum(const void * iph,unsigned int ihl)660 static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
661 {
662 	return (__sum16)~do_csum(iph, ihl * 4);
663 }
664 
665 /*
666  * Fold a partial checksum
667  * This function code has been taken from
668  * Linux kernel include/asm-generic/checksum.h
669  */
csum_fold(__wsum csum)670 static inline __sum16 csum_fold(__wsum csum)
671 {
672 	u32 sum = (u32)csum;
673 
674 	sum = (sum & 0xffff) + (sum >> 16);
675 	sum = (sum & 0xffff) + (sum >> 16);
676 	return (__sum16)~sum;
677 }
678 
679 /*
680  * This function code has been taken from
681  * Linux kernel lib/checksum.c
682  */
from64to32(u64 x)683 static inline u32 from64to32(u64 x)
684 {
685 	/* add up 32-bit and 32-bit for 32+c bit */
686 	x = (x & 0xffffffff) + (x >> 32);
687 	/* add up carry.. */
688 	x = (x & 0xffffffff) + (x >> 32);
689 	return (u32)x;
690 }
691 
692 __wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
693 			  __u32 len, __u8 proto, __wsum sum);
694 
695 /*
696  * This function code has been taken from
697  * Linux kernel lib/checksum.c
698  */
csum_tcpudp_nofold(__be32 saddr,__be32 daddr,__u32 len,__u8 proto,__wsum sum)699 __wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
700 			  __u32 len, __u8 proto, __wsum sum)
701 {
702 	unsigned long long s = (u32)sum;
703 
704 	s += (u32)saddr;
705 	s += (u32)daddr;
706 #ifdef __BIG_ENDIAN__
707 	s += proto + len;
708 #else
709 	s += (proto + len) << 8;
710 #endif
711 	return (__wsum)from64to32(s);
712 }
713 
714 /*
715  * This function has been taken from
716  * Linux kernel include/asm-generic/checksum.h
717  */
718 static inline __sum16
csum_tcpudp_magic(__be32 saddr,__be32 daddr,__u32 len,__u8 proto,__wsum sum)719 csum_tcpudp_magic(__be32 saddr, __be32 daddr, __u32 len,
720 		  __u8 proto, __wsum sum)
721 {
722 	return csum_fold(csum_tcpudp_nofold(saddr, daddr, len, proto, sum));
723 }
724 
udp_csum(u32 saddr,u32 daddr,u32 len,u8 proto,u16 * udp_pkt)725 static inline u16 udp_csum(u32 saddr, u32 daddr, u32 len,
726 			   u8 proto, u16 *udp_pkt)
727 {
728 	u32 csum = 0;
729 	u32 cnt = 0;
730 
731 	/* udp hdr and data */
732 	for (; cnt < len; cnt += 2)
733 		csum += udp_pkt[cnt >> 1];
734 
735 	return csum_tcpudp_magic(saddr, daddr, len, proto, csum);
736 }
737 
738 #define ETH_FCS_SIZE 4
739 
740 #define PKT_HDR_SIZE (sizeof(struct ethhdr) + sizeof(struct iphdr) + \
741 		      sizeof(struct udphdr))
742 
743 #define PKT_SIZE		(opt_pkt_size - ETH_FCS_SIZE)
744 #define IP_PKT_SIZE		(PKT_SIZE - sizeof(struct ethhdr))
745 #define UDP_PKT_SIZE		(IP_PKT_SIZE - sizeof(struct iphdr))
746 #define UDP_PKT_DATA_SIZE	(UDP_PKT_SIZE - sizeof(struct udphdr))
747 
748 static u8 pkt_data[XSK_UMEM__DEFAULT_FRAME_SIZE];
749 
gen_eth_hdr_data(void)750 static void gen_eth_hdr_data(void)
751 {
752 	struct udphdr *udp_hdr = (struct udphdr *)(pkt_data +
753 						   sizeof(struct ethhdr) +
754 						   sizeof(struct iphdr));
755 	struct iphdr *ip_hdr = (struct iphdr *)(pkt_data +
756 						sizeof(struct ethhdr));
757 	struct ethhdr *eth_hdr = (struct ethhdr *)pkt_data;
758 
759 	/* ethernet header */
760 	memcpy(eth_hdr->h_dest, "\x3c\xfd\xfe\x9e\x7f\x71", ETH_ALEN);
761 	memcpy(eth_hdr->h_source, "\xec\xb1\xd7\x98\x3a\xc0", ETH_ALEN);
762 	eth_hdr->h_proto = htons(ETH_P_IP);
763 
764 	/* IP header */
765 	ip_hdr->version = IPVERSION;
766 	ip_hdr->ihl = 0x5; /* 20 byte header */
767 	ip_hdr->tos = 0x0;
768 	ip_hdr->tot_len = htons(IP_PKT_SIZE);
769 	ip_hdr->id = 0;
770 	ip_hdr->frag_off = 0;
771 	ip_hdr->ttl = IPDEFTTL;
772 	ip_hdr->protocol = IPPROTO_UDP;
773 	ip_hdr->saddr = htonl(0x0a0a0a10);
774 	ip_hdr->daddr = htonl(0x0a0a0a20);
775 
776 	/* IP header checksum */
777 	ip_hdr->check = 0;
778 	ip_hdr->check = ip_fast_csum((const void *)ip_hdr, ip_hdr->ihl);
779 
780 	/* UDP header */
781 	udp_hdr->source = htons(0x1000);
782 	udp_hdr->dest = htons(0x1000);
783 	udp_hdr->len = htons(UDP_PKT_SIZE);
784 
785 	/* UDP data */
786 	memset32_htonl(pkt_data + PKT_HDR_SIZE, opt_pkt_fill_pattern,
787 		       UDP_PKT_DATA_SIZE);
788 
789 	/* UDP header checksum */
790 	udp_hdr->check = 0;
791 	udp_hdr->check = udp_csum(ip_hdr->saddr, ip_hdr->daddr, UDP_PKT_SIZE,
792 				  IPPROTO_UDP, (u16 *)udp_hdr);
793 }
794 
gen_eth_frame(struct xsk_umem_info * umem,u64 addr)795 static void gen_eth_frame(struct xsk_umem_info *umem, u64 addr)
796 {
797 	memcpy(xsk_umem__get_data(umem->buffer, addr), pkt_data,
798 	       PKT_SIZE);
799 }
800 
xsk_configure_umem(void * buffer,u64 size)801 static struct xsk_umem_info *xsk_configure_umem(void *buffer, u64 size)
802 {
803 	struct xsk_umem_info *umem;
804 	struct xsk_umem_config cfg = {
805 		/* We recommend that you set the fill ring size >= HW RX ring size +
806 		 * AF_XDP RX ring size. Make sure you fill up the fill ring
807 		 * with buffers at regular intervals, and you will with this setting
808 		 * avoid allocation failures in the driver. These are usually quite
809 		 * expensive since drivers have not been written to assume that
810 		 * allocation failures are common. For regular sockets, kernel
811 		 * allocated memory is used that only runs out in OOM situations
812 		 * that should be rare.
813 		 */
814 		.fill_size = XSK_RING_PROD__DEFAULT_NUM_DESCS * 2,
815 		.comp_size = XSK_RING_CONS__DEFAULT_NUM_DESCS,
816 		.frame_size = opt_xsk_frame_size,
817 		.frame_headroom = XSK_UMEM__DEFAULT_FRAME_HEADROOM,
818 		.flags = opt_umem_flags
819 	};
820 	int ret;
821 
822 	umem = calloc(1, sizeof(*umem));
823 	if (!umem)
824 		exit_with_error(errno);
825 
826 	ret = xsk_umem__create(&umem->umem, buffer, size, &umem->fq, &umem->cq,
827 			       &cfg);
828 	if (ret)
829 		exit_with_error(-ret);
830 
831 	umem->buffer = buffer;
832 	return umem;
833 }
834 
xsk_populate_fill_ring(struct xsk_umem_info * umem)835 static void xsk_populate_fill_ring(struct xsk_umem_info *umem)
836 {
837 	int ret, i;
838 	u32 idx;
839 
840 	ret = xsk_ring_prod__reserve(&umem->fq,
841 				     XSK_RING_PROD__DEFAULT_NUM_DESCS * 2, &idx);
842 	if (ret != XSK_RING_PROD__DEFAULT_NUM_DESCS * 2)
843 		exit_with_error(-ret);
844 	for (i = 0; i < XSK_RING_PROD__DEFAULT_NUM_DESCS * 2; i++)
845 		*xsk_ring_prod__fill_addr(&umem->fq, idx++) =
846 			i * opt_xsk_frame_size;
847 	xsk_ring_prod__submit(&umem->fq, XSK_RING_PROD__DEFAULT_NUM_DESCS * 2);
848 }
849 
xsk_configure_socket(struct xsk_umem_info * umem,bool rx,bool tx)850 static struct xsk_socket_info *xsk_configure_socket(struct xsk_umem_info *umem,
851 						    bool rx, bool tx)
852 {
853 	struct xsk_socket_config cfg;
854 	struct xsk_socket_info *xsk;
855 	struct xsk_ring_cons *rxr;
856 	struct xsk_ring_prod *txr;
857 	int ret;
858 
859 	xsk = calloc(1, sizeof(*xsk));
860 	if (!xsk)
861 		exit_with_error(errno);
862 
863 	xsk->umem = umem;
864 	cfg.rx_size = XSK_RING_CONS__DEFAULT_NUM_DESCS;
865 	cfg.tx_size = XSK_RING_PROD__DEFAULT_NUM_DESCS;
866 	if (opt_num_xsks > 1 || opt_reduced_cap)
867 		cfg.libbpf_flags = XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD;
868 	else
869 		cfg.libbpf_flags = 0;
870 	cfg.xdp_flags = opt_xdp_flags;
871 	cfg.bind_flags = opt_xdp_bind_flags;
872 
873 	rxr = rx ? &xsk->rx : NULL;
874 	txr = tx ? &xsk->tx : NULL;
875 	ret = xsk_socket__create(&xsk->xsk, opt_if, opt_queue, umem->umem,
876 				 rxr, txr, &cfg);
877 	if (ret)
878 		exit_with_error(-ret);
879 
880 	ret = bpf_get_link_xdp_id(opt_ifindex, &prog_id, opt_xdp_flags);
881 	if (ret)
882 		exit_with_error(-ret);
883 
884 	xsk->app_stats.rx_empty_polls = 0;
885 	xsk->app_stats.fill_fail_polls = 0;
886 	xsk->app_stats.copy_tx_sendtos = 0;
887 	xsk->app_stats.tx_wakeup_sendtos = 0;
888 	xsk->app_stats.opt_polls = 0;
889 	xsk->app_stats.prev_rx_empty_polls = 0;
890 	xsk->app_stats.prev_fill_fail_polls = 0;
891 	xsk->app_stats.prev_copy_tx_sendtos = 0;
892 	xsk->app_stats.prev_tx_wakeup_sendtos = 0;
893 	xsk->app_stats.prev_opt_polls = 0;
894 
895 	return xsk;
896 }
897 
898 static struct option long_options[] = {
899 	{"rxdrop", no_argument, 0, 'r'},
900 	{"txonly", no_argument, 0, 't'},
901 	{"l2fwd", no_argument, 0, 'l'},
902 	{"interface", required_argument, 0, 'i'},
903 	{"queue", required_argument, 0, 'q'},
904 	{"poll", no_argument, 0, 'p'},
905 	{"xdp-skb", no_argument, 0, 'S'},
906 	{"xdp-native", no_argument, 0, 'N'},
907 	{"interval", required_argument, 0, 'n'},
908 	{"zero-copy", no_argument, 0, 'z'},
909 	{"copy", no_argument, 0, 'c'},
910 	{"frame-size", required_argument, 0, 'f'},
911 	{"no-need-wakeup", no_argument, 0, 'm'},
912 	{"unaligned", no_argument, 0, 'u'},
913 	{"shared-umem", no_argument, 0, 'M'},
914 	{"force", no_argument, 0, 'F'},
915 	{"duration", required_argument, 0, 'd'},
916 	{"batch-size", required_argument, 0, 'b'},
917 	{"tx-pkt-count", required_argument, 0, 'C'},
918 	{"tx-pkt-size", required_argument, 0, 's'},
919 	{"tx-pkt-pattern", required_argument, 0, 'P'},
920 	{"extra-stats", no_argument, 0, 'x'},
921 	{"quiet", no_argument, 0, 'Q'},
922 	{"app-stats", no_argument, 0, 'a'},
923 	{"irq-string", no_argument, 0, 'I'},
924 	{"busy-poll", no_argument, 0, 'B'},
925 	{"reduce-cap", no_argument, 0, 'R'},
926 	{0, 0, 0, 0}
927 };
928 
usage(const char * prog)929 static void usage(const char *prog)
930 {
931 	const char *str =
932 		"  Usage: %s [OPTIONS]\n"
933 		"  Options:\n"
934 		"  -r, --rxdrop		Discard all incoming packets (default)\n"
935 		"  -t, --txonly		Only send packets\n"
936 		"  -l, --l2fwd		MAC swap L2 forwarding\n"
937 		"  -i, --interface=n	Run on interface n\n"
938 		"  -q, --queue=n	Use queue n (default 0)\n"
939 		"  -p, --poll		Use poll syscall\n"
940 		"  -S, --xdp-skb=n	Use XDP skb-mod\n"
941 		"  -N, --xdp-native=n	Enforce XDP native mode\n"
942 		"  -n, --interval=n	Specify statistics update interval (default 1 sec).\n"
943 		"  -z, --zero-copy      Force zero-copy mode.\n"
944 		"  -c, --copy           Force copy mode.\n"
945 		"  -m, --no-need-wakeup Turn off use of driver need wakeup flag.\n"
946 		"  -f, --frame-size=n   Set the frame size (must be a power of two in aligned mode, default is %d).\n"
947 		"  -u, --unaligned	Enable unaligned chunk placement\n"
948 		"  -M, --shared-umem	Enable XDP_SHARED_UMEM (cannot be used with -R)\n"
949 		"  -F, --force		Force loading the XDP prog\n"
950 		"  -d, --duration=n	Duration in secs to run command.\n"
951 		"			Default: forever.\n"
952 		"  -b, --batch-size=n	Batch size for sending or receiving\n"
953 		"			packets. Default: %d\n"
954 		"  -C, --tx-pkt-count=n	Number of packets to send.\n"
955 		"			Default: Continuous packets.\n"
956 		"  -s, --tx-pkt-size=n	Transmit packet size.\n"
957 		"			(Default: %d bytes)\n"
958 		"			Min size: %d, Max size %d.\n"
959 		"  -P, --tx-pkt-pattern=nPacket fill pattern. Default: 0x%x\n"
960 		"  -x, --extra-stats	Display extra statistics.\n"
961 		"  -Q, --quiet          Do not display any stats.\n"
962 		"  -a, --app-stats	Display application (syscall) statistics.\n"
963 		"  -I, --irq-string	Display driver interrupt statistics for interface associated with irq-string.\n"
964 		"  -B, --busy-poll      Busy poll.\n"
965 		"  -R, --reduce-cap	Use reduced capabilities (cannot be used with -M)\n"
966 		"\n";
967 	fprintf(stderr, str, prog, XSK_UMEM__DEFAULT_FRAME_SIZE,
968 		opt_batch_size, MIN_PKT_SIZE, MIN_PKT_SIZE,
969 		XSK_UMEM__DEFAULT_FRAME_SIZE, opt_pkt_fill_pattern);
970 
971 	exit(EXIT_FAILURE);
972 }
973 
parse_command_line(int argc,char ** argv)974 static void parse_command_line(int argc, char **argv)
975 {
976 	int option_index, c;
977 
978 	opterr = 0;
979 
980 	for (;;) {
981 		c = getopt_long(argc, argv, "Frtli:q:pSNn:czf:muMd:b:C:s:P:xQaI:BR",
982 				long_options, &option_index);
983 		if (c == -1)
984 			break;
985 
986 		switch (c) {
987 		case 'r':
988 			opt_bench = BENCH_RXDROP;
989 			break;
990 		case 't':
991 			opt_bench = BENCH_TXONLY;
992 			break;
993 		case 'l':
994 			opt_bench = BENCH_L2FWD;
995 			break;
996 		case 'i':
997 			opt_if = optarg;
998 			break;
999 		case 'q':
1000 			opt_queue = atoi(optarg);
1001 			break;
1002 		case 'p':
1003 			opt_poll = 1;
1004 			break;
1005 		case 'S':
1006 			opt_xdp_flags |= XDP_FLAGS_SKB_MODE;
1007 			opt_xdp_bind_flags |= XDP_COPY;
1008 			break;
1009 		case 'N':
1010 			/* default, set below */
1011 			break;
1012 		case 'n':
1013 			opt_interval = atoi(optarg);
1014 			break;
1015 		case 'z':
1016 			opt_xdp_bind_flags |= XDP_ZEROCOPY;
1017 			break;
1018 		case 'c':
1019 			opt_xdp_bind_flags |= XDP_COPY;
1020 			break;
1021 		case 'u':
1022 			opt_umem_flags |= XDP_UMEM_UNALIGNED_CHUNK_FLAG;
1023 			opt_unaligned_chunks = 1;
1024 			opt_mmap_flags = MAP_HUGETLB;
1025 			break;
1026 		case 'F':
1027 			opt_xdp_flags &= ~XDP_FLAGS_UPDATE_IF_NOEXIST;
1028 			break;
1029 		case 'f':
1030 			opt_xsk_frame_size = atoi(optarg);
1031 			break;
1032 		case 'm':
1033 			opt_need_wakeup = false;
1034 			opt_xdp_bind_flags &= ~XDP_USE_NEED_WAKEUP;
1035 			break;
1036 		case 'M':
1037 			opt_num_xsks = MAX_SOCKS;
1038 			break;
1039 		case 'd':
1040 			opt_duration = atoi(optarg);
1041 			opt_duration *= 1000000000;
1042 			break;
1043 		case 'b':
1044 			opt_batch_size = atoi(optarg);
1045 			break;
1046 		case 'C':
1047 			opt_pkt_count = atoi(optarg);
1048 			break;
1049 		case 's':
1050 			opt_pkt_size = atoi(optarg);
1051 			if (opt_pkt_size > (XSK_UMEM__DEFAULT_FRAME_SIZE) ||
1052 			    opt_pkt_size < MIN_PKT_SIZE) {
1053 				fprintf(stderr,
1054 					"ERROR: Invalid frame size %d\n",
1055 					opt_pkt_size);
1056 				usage(basename(argv[0]));
1057 			}
1058 			break;
1059 		case 'P':
1060 			opt_pkt_fill_pattern = strtol(optarg, NULL, 16);
1061 			break;
1062 		case 'x':
1063 			opt_extra_stats = 1;
1064 			break;
1065 		case 'Q':
1066 			opt_quiet = 1;
1067 			break;
1068 		case 'a':
1069 			opt_app_stats = 1;
1070 			break;
1071 		case 'I':
1072 			opt_irq_str = optarg;
1073 			if (get_interrupt_number())
1074 				irqs_at_init = get_irqs();
1075 			if (irqs_at_init < 0) {
1076 				fprintf(stderr, "ERROR: Failed to get irqs for %s\n", opt_irq_str);
1077 				usage(basename(argv[0]));
1078 			}
1079 			break;
1080 		case 'B':
1081 			opt_busy_poll = 1;
1082 			break;
1083 		case 'R':
1084 			opt_reduced_cap = true;
1085 			break;
1086 		default:
1087 			usage(basename(argv[0]));
1088 		}
1089 	}
1090 
1091 	if (!(opt_xdp_flags & XDP_FLAGS_SKB_MODE))
1092 		opt_xdp_flags |= XDP_FLAGS_DRV_MODE;
1093 
1094 	opt_ifindex = if_nametoindex(opt_if);
1095 	if (!opt_ifindex) {
1096 		fprintf(stderr, "ERROR: interface \"%s\" does not exist\n",
1097 			opt_if);
1098 		usage(basename(argv[0]));
1099 	}
1100 
1101 	if ((opt_xsk_frame_size & (opt_xsk_frame_size - 1)) &&
1102 	    !opt_unaligned_chunks) {
1103 		fprintf(stderr, "--frame-size=%d is not a power of two\n",
1104 			opt_xsk_frame_size);
1105 		usage(basename(argv[0]));
1106 	}
1107 
1108 	if (opt_reduced_cap && opt_num_xsks > 1) {
1109 		fprintf(stderr, "ERROR: -M and -R cannot be used together\n");
1110 		usage(basename(argv[0]));
1111 	}
1112 }
1113 
kick_tx(struct xsk_socket_info * xsk)1114 static void kick_tx(struct xsk_socket_info *xsk)
1115 {
1116 	int ret;
1117 
1118 	ret = sendto(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, 0);
1119 	if (ret >= 0 || errno == ENOBUFS || errno == EAGAIN ||
1120 	    errno == EBUSY || errno == ENETDOWN)
1121 		return;
1122 	exit_with_error(errno);
1123 }
1124 
complete_tx_l2fwd(struct xsk_socket_info * xsk)1125 static inline void complete_tx_l2fwd(struct xsk_socket_info *xsk)
1126 {
1127 	struct xsk_umem_info *umem = xsk->umem;
1128 	u32 idx_cq = 0, idx_fq = 0;
1129 	unsigned int rcvd;
1130 	size_t ndescs;
1131 
1132 	if (!xsk->outstanding_tx)
1133 		return;
1134 
1135 	/* In copy mode, Tx is driven by a syscall so we need to use e.g. sendto() to
1136 	 * really send the packets. In zero-copy mode we do not have to do this, since Tx
1137 	 * is driven by the NAPI loop. So as an optimization, we do not have to call
1138 	 * sendto() all the time in zero-copy mode for l2fwd.
1139 	 */
1140 	if (opt_xdp_bind_flags & XDP_COPY) {
1141 		xsk->app_stats.copy_tx_sendtos++;
1142 		kick_tx(xsk);
1143 	}
1144 
1145 	ndescs = (xsk->outstanding_tx > opt_batch_size) ? opt_batch_size :
1146 		xsk->outstanding_tx;
1147 
1148 	/* re-add completed Tx buffers */
1149 	rcvd = xsk_ring_cons__peek(&umem->cq, ndescs, &idx_cq);
1150 	if (rcvd > 0) {
1151 		unsigned int i;
1152 		int ret;
1153 
1154 		ret = xsk_ring_prod__reserve(&umem->fq, rcvd, &idx_fq);
1155 		while (ret != rcvd) {
1156 			if (ret < 0)
1157 				exit_with_error(-ret);
1158 			if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&umem->fq)) {
1159 				xsk->app_stats.fill_fail_polls++;
1160 				recvfrom(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL,
1161 					 NULL);
1162 			}
1163 			ret = xsk_ring_prod__reserve(&umem->fq, rcvd, &idx_fq);
1164 		}
1165 
1166 		for (i = 0; i < rcvd; i++)
1167 			*xsk_ring_prod__fill_addr(&umem->fq, idx_fq++) =
1168 				*xsk_ring_cons__comp_addr(&umem->cq, idx_cq++);
1169 
1170 		xsk_ring_prod__submit(&xsk->umem->fq, rcvd);
1171 		xsk_ring_cons__release(&xsk->umem->cq, rcvd);
1172 		xsk->outstanding_tx -= rcvd;
1173 	}
1174 }
1175 
complete_tx_only(struct xsk_socket_info * xsk,int batch_size)1176 static inline void complete_tx_only(struct xsk_socket_info *xsk,
1177 				    int batch_size)
1178 {
1179 	unsigned int rcvd;
1180 	u32 idx;
1181 
1182 	if (!xsk->outstanding_tx)
1183 		return;
1184 
1185 	if (!opt_need_wakeup || xsk_ring_prod__needs_wakeup(&xsk->tx)) {
1186 		xsk->app_stats.tx_wakeup_sendtos++;
1187 		kick_tx(xsk);
1188 	}
1189 
1190 	rcvd = xsk_ring_cons__peek(&xsk->umem->cq, batch_size, &idx);
1191 	if (rcvd > 0) {
1192 		xsk_ring_cons__release(&xsk->umem->cq, rcvd);
1193 		xsk->outstanding_tx -= rcvd;
1194 	}
1195 }
1196 
rx_drop(struct xsk_socket_info * xsk)1197 static void rx_drop(struct xsk_socket_info *xsk)
1198 {
1199 	unsigned int rcvd, i;
1200 	u32 idx_rx = 0, idx_fq = 0;
1201 	int ret;
1202 
1203 	rcvd = xsk_ring_cons__peek(&xsk->rx, opt_batch_size, &idx_rx);
1204 	if (!rcvd) {
1205 		if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&xsk->umem->fq)) {
1206 			xsk->app_stats.rx_empty_polls++;
1207 			recvfrom(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, NULL);
1208 		}
1209 		return;
1210 	}
1211 
1212 	ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
1213 	while (ret != rcvd) {
1214 		if (ret < 0)
1215 			exit_with_error(-ret);
1216 		if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&xsk->umem->fq)) {
1217 			xsk->app_stats.fill_fail_polls++;
1218 			recvfrom(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, NULL);
1219 		}
1220 		ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
1221 	}
1222 
1223 	for (i = 0; i < rcvd; i++) {
1224 		u64 addr = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx)->addr;
1225 		u32 len = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx++)->len;
1226 		u64 orig = xsk_umem__extract_addr(addr);
1227 
1228 		addr = xsk_umem__add_offset_to_addr(addr);
1229 		char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
1230 
1231 		hex_dump(pkt, len, addr);
1232 		*xsk_ring_prod__fill_addr(&xsk->umem->fq, idx_fq++) = orig;
1233 	}
1234 
1235 	xsk_ring_prod__submit(&xsk->umem->fq, rcvd);
1236 	xsk_ring_cons__release(&xsk->rx, rcvd);
1237 	xsk->ring_stats.rx_npkts += rcvd;
1238 }
1239 
rx_drop_all(void)1240 static void rx_drop_all(void)
1241 {
1242 	struct pollfd fds[MAX_SOCKS] = {};
1243 	int i, ret;
1244 
1245 	for (i = 0; i < num_socks; i++) {
1246 		fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
1247 		fds[i].events = POLLIN;
1248 	}
1249 
1250 	for (;;) {
1251 		if (opt_poll) {
1252 			for (i = 0; i < num_socks; i++)
1253 				xsks[i]->app_stats.opt_polls++;
1254 			ret = poll(fds, num_socks, opt_timeout);
1255 			if (ret <= 0)
1256 				continue;
1257 		}
1258 
1259 		for (i = 0; i < num_socks; i++)
1260 			rx_drop(xsks[i]);
1261 
1262 		if (benchmark_done)
1263 			break;
1264 	}
1265 }
1266 
tx_only(struct xsk_socket_info * xsk,u32 * frame_nb,int batch_size)1267 static void tx_only(struct xsk_socket_info *xsk, u32 *frame_nb, int batch_size)
1268 {
1269 	u32 idx;
1270 	unsigned int i;
1271 
1272 	while (xsk_ring_prod__reserve(&xsk->tx, batch_size, &idx) <
1273 				      batch_size) {
1274 		complete_tx_only(xsk, batch_size);
1275 		if (benchmark_done)
1276 			return;
1277 	}
1278 
1279 	for (i = 0; i < batch_size; i++) {
1280 		struct xdp_desc *tx_desc = xsk_ring_prod__tx_desc(&xsk->tx,
1281 								  idx + i);
1282 		tx_desc->addr = (*frame_nb + i) * opt_xsk_frame_size;
1283 		tx_desc->len = PKT_SIZE;
1284 	}
1285 
1286 	xsk_ring_prod__submit(&xsk->tx, batch_size);
1287 	xsk->ring_stats.tx_npkts += batch_size;
1288 	xsk->outstanding_tx += batch_size;
1289 	*frame_nb += batch_size;
1290 	*frame_nb %= NUM_FRAMES;
1291 	complete_tx_only(xsk, batch_size);
1292 }
1293 
get_batch_size(int pkt_cnt)1294 static inline int get_batch_size(int pkt_cnt)
1295 {
1296 	if (!opt_pkt_count)
1297 		return opt_batch_size;
1298 
1299 	if (pkt_cnt + opt_batch_size <= opt_pkt_count)
1300 		return opt_batch_size;
1301 
1302 	return opt_pkt_count - pkt_cnt;
1303 }
1304 
complete_tx_only_all(void)1305 static void complete_tx_only_all(void)
1306 {
1307 	bool pending;
1308 	int i;
1309 
1310 	do {
1311 		pending = false;
1312 		for (i = 0; i < num_socks; i++) {
1313 			if (xsks[i]->outstanding_tx) {
1314 				complete_tx_only(xsks[i], opt_batch_size);
1315 				pending = !!xsks[i]->outstanding_tx;
1316 			}
1317 		}
1318 	} while (pending);
1319 }
1320 
tx_only_all(void)1321 static void tx_only_all(void)
1322 {
1323 	struct pollfd fds[MAX_SOCKS] = {};
1324 	u32 frame_nb[MAX_SOCKS] = {};
1325 	int pkt_cnt = 0;
1326 	int i, ret;
1327 
1328 	for (i = 0; i < num_socks; i++) {
1329 		fds[0].fd = xsk_socket__fd(xsks[i]->xsk);
1330 		fds[0].events = POLLOUT;
1331 	}
1332 
1333 	while ((opt_pkt_count && pkt_cnt < opt_pkt_count) || !opt_pkt_count) {
1334 		int batch_size = get_batch_size(pkt_cnt);
1335 
1336 		if (opt_poll) {
1337 			for (i = 0; i < num_socks; i++)
1338 				xsks[i]->app_stats.opt_polls++;
1339 			ret = poll(fds, num_socks, opt_timeout);
1340 			if (ret <= 0)
1341 				continue;
1342 
1343 			if (!(fds[0].revents & POLLOUT))
1344 				continue;
1345 		}
1346 
1347 		for (i = 0; i < num_socks; i++)
1348 			tx_only(xsks[i], &frame_nb[i], batch_size);
1349 
1350 		pkt_cnt += batch_size;
1351 
1352 		if (benchmark_done)
1353 			break;
1354 	}
1355 
1356 	if (opt_pkt_count)
1357 		complete_tx_only_all();
1358 }
1359 
l2fwd(struct xsk_socket_info * xsk)1360 static void l2fwd(struct xsk_socket_info *xsk)
1361 {
1362 	unsigned int rcvd, i;
1363 	u32 idx_rx = 0, idx_tx = 0;
1364 	int ret;
1365 
1366 	complete_tx_l2fwd(xsk);
1367 
1368 	rcvd = xsk_ring_cons__peek(&xsk->rx, opt_batch_size, &idx_rx);
1369 	if (!rcvd) {
1370 		if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&xsk->umem->fq)) {
1371 			xsk->app_stats.rx_empty_polls++;
1372 			recvfrom(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, NULL);
1373 		}
1374 		return;
1375 	}
1376 	xsk->ring_stats.rx_npkts += rcvd;
1377 
1378 	ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
1379 	while (ret != rcvd) {
1380 		if (ret < 0)
1381 			exit_with_error(-ret);
1382 		complete_tx_l2fwd(xsk);
1383 		if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&xsk->tx)) {
1384 			xsk->app_stats.tx_wakeup_sendtos++;
1385 			kick_tx(xsk);
1386 		}
1387 		ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
1388 	}
1389 
1390 	for (i = 0; i < rcvd; i++) {
1391 		u64 addr = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx)->addr;
1392 		u32 len = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx++)->len;
1393 		u64 orig = addr;
1394 
1395 		addr = xsk_umem__add_offset_to_addr(addr);
1396 		char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
1397 
1398 		swap_mac_addresses(pkt);
1399 
1400 		hex_dump(pkt, len, addr);
1401 		xsk_ring_prod__tx_desc(&xsk->tx, idx_tx)->addr = orig;
1402 		xsk_ring_prod__tx_desc(&xsk->tx, idx_tx++)->len = len;
1403 	}
1404 
1405 	xsk_ring_prod__submit(&xsk->tx, rcvd);
1406 	xsk_ring_cons__release(&xsk->rx, rcvd);
1407 
1408 	xsk->ring_stats.tx_npkts += rcvd;
1409 	xsk->outstanding_tx += rcvd;
1410 }
1411 
l2fwd_all(void)1412 static void l2fwd_all(void)
1413 {
1414 	struct pollfd fds[MAX_SOCKS] = {};
1415 	int i, ret;
1416 
1417 	for (;;) {
1418 		if (opt_poll) {
1419 			for (i = 0; i < num_socks; i++) {
1420 				fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
1421 				fds[i].events = POLLOUT | POLLIN;
1422 				xsks[i]->app_stats.opt_polls++;
1423 			}
1424 			ret = poll(fds, num_socks, opt_timeout);
1425 			if (ret <= 0)
1426 				continue;
1427 		}
1428 
1429 		for (i = 0; i < num_socks; i++)
1430 			l2fwd(xsks[i]);
1431 
1432 		if (benchmark_done)
1433 			break;
1434 	}
1435 }
1436 
load_xdp_program(char ** argv,struct bpf_object ** obj)1437 static void load_xdp_program(char **argv, struct bpf_object **obj)
1438 {
1439 	struct bpf_prog_load_attr prog_load_attr = {
1440 		.prog_type      = BPF_PROG_TYPE_XDP,
1441 	};
1442 	char xdp_filename[256];
1443 	int prog_fd;
1444 
1445 	snprintf(xdp_filename, sizeof(xdp_filename), "%s_kern.o", argv[0]);
1446 	prog_load_attr.file = xdp_filename;
1447 
1448 	if (bpf_prog_load_xattr(&prog_load_attr, obj, &prog_fd))
1449 		exit(EXIT_FAILURE);
1450 	if (prog_fd < 0) {
1451 		fprintf(stderr, "ERROR: no program found: %s\n",
1452 			strerror(prog_fd));
1453 		exit(EXIT_FAILURE);
1454 	}
1455 
1456 	if (bpf_set_link_xdp_fd(opt_ifindex, prog_fd, opt_xdp_flags) < 0) {
1457 		fprintf(stderr, "ERROR: link set xdp fd failed\n");
1458 		exit(EXIT_FAILURE);
1459 	}
1460 }
1461 
enter_xsks_into_map(struct bpf_object * obj)1462 static void enter_xsks_into_map(struct bpf_object *obj)
1463 {
1464 	struct bpf_map *map;
1465 	int i, xsks_map;
1466 
1467 	map = bpf_object__find_map_by_name(obj, "xsks_map");
1468 	xsks_map = bpf_map__fd(map);
1469 	if (xsks_map < 0) {
1470 		fprintf(stderr, "ERROR: no xsks map found: %s\n",
1471 			strerror(xsks_map));
1472 			exit(EXIT_FAILURE);
1473 	}
1474 
1475 	for (i = 0; i < num_socks; i++) {
1476 		int fd = xsk_socket__fd(xsks[i]->xsk);
1477 		int key, ret;
1478 
1479 		key = i;
1480 		ret = bpf_map_update_elem(xsks_map, &key, &fd, 0);
1481 		if (ret) {
1482 			fprintf(stderr, "ERROR: bpf_map_update_elem %d\n", i);
1483 			exit(EXIT_FAILURE);
1484 		}
1485 	}
1486 }
1487 
apply_setsockopt(struct xsk_socket_info * xsk)1488 static void apply_setsockopt(struct xsk_socket_info *xsk)
1489 {
1490 	int sock_opt;
1491 
1492 	if (!opt_busy_poll)
1493 		return;
1494 
1495 	sock_opt = 1;
1496 	if (setsockopt(xsk_socket__fd(xsk->xsk), SOL_SOCKET, SO_PREFER_BUSY_POLL,
1497 		       (void *)&sock_opt, sizeof(sock_opt)) < 0)
1498 		exit_with_error(errno);
1499 
1500 	sock_opt = 20;
1501 	if (setsockopt(xsk_socket__fd(xsk->xsk), SOL_SOCKET, SO_BUSY_POLL,
1502 		       (void *)&sock_opt, sizeof(sock_opt)) < 0)
1503 		exit_with_error(errno);
1504 
1505 	sock_opt = opt_batch_size;
1506 	if (setsockopt(xsk_socket__fd(xsk->xsk), SOL_SOCKET, SO_BUSY_POLL_BUDGET,
1507 		       (void *)&sock_opt, sizeof(sock_opt)) < 0)
1508 		exit_with_error(errno);
1509 }
1510 
recv_xsks_map_fd_from_ctrl_node(int sock,int * _fd)1511 static int recv_xsks_map_fd_from_ctrl_node(int sock, int *_fd)
1512 {
1513 	char cms[CMSG_SPACE(sizeof(int))];
1514 	struct cmsghdr *cmsg;
1515 	struct msghdr msg;
1516 	struct iovec iov;
1517 	int value;
1518 	int len;
1519 
1520 	iov.iov_base = &value;
1521 	iov.iov_len = sizeof(int);
1522 
1523 	msg.msg_name = 0;
1524 	msg.msg_namelen = 0;
1525 	msg.msg_iov = &iov;
1526 	msg.msg_iovlen = 1;
1527 	msg.msg_flags = 0;
1528 	msg.msg_control = (caddr_t)cms;
1529 	msg.msg_controllen = sizeof(cms);
1530 
1531 	len = recvmsg(sock, &msg, 0);
1532 
1533 	if (len < 0) {
1534 		fprintf(stderr, "Recvmsg failed length incorrect.\n");
1535 		return -EINVAL;
1536 	}
1537 
1538 	if (len == 0) {
1539 		fprintf(stderr, "Recvmsg failed no data\n");
1540 		return -EINVAL;
1541 	}
1542 
1543 	cmsg = CMSG_FIRSTHDR(&msg);
1544 	*_fd = *(int *)CMSG_DATA(cmsg);
1545 
1546 	return 0;
1547 }
1548 
1549 static int
recv_xsks_map_fd(int * xsks_map_fd)1550 recv_xsks_map_fd(int *xsks_map_fd)
1551 {
1552 	struct sockaddr_un server;
1553 	int err;
1554 
1555 	sock = socket(AF_UNIX, SOCK_STREAM, 0);
1556 	if (sock < 0) {
1557 		fprintf(stderr, "Error opening socket stream: %s", strerror(errno));
1558 		return errno;
1559 	}
1560 
1561 	server.sun_family = AF_UNIX;
1562 	strcpy(server.sun_path, SOCKET_NAME);
1563 
1564 	if (connect(sock, (struct sockaddr *)&server, sizeof(struct sockaddr_un)) < 0) {
1565 		close(sock);
1566 		fprintf(stderr, "Error connecting stream socket: %s", strerror(errno));
1567 		return errno;
1568 	}
1569 
1570 	err = recv_xsks_map_fd_from_ctrl_node(sock, xsks_map_fd);
1571 	if (err) {
1572 		fprintf(stderr, "Error %d receiving fd\n", err);
1573 		return err;
1574 	}
1575 	return 0;
1576 }
1577 
main(int argc,char ** argv)1578 int main(int argc, char **argv)
1579 {
1580 	struct __user_cap_header_struct hdr = { _LINUX_CAPABILITY_VERSION_3, 0 };
1581 	struct __user_cap_data_struct data[2] = { { 0 } };
1582 	struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
1583 	bool rx = false, tx = false;
1584 	struct xsk_umem_info *umem;
1585 	struct bpf_object *obj;
1586 	int xsks_map_fd = 0;
1587 	pthread_t pt;
1588 	int i, ret;
1589 	void *bufs;
1590 
1591 	parse_command_line(argc, argv);
1592 
1593 	if (opt_reduced_cap) {
1594 		if (capget(&hdr, data)  < 0)
1595 			fprintf(stderr, "Error getting capabilities\n");
1596 
1597 		data->effective &= CAP_TO_MASK(CAP_NET_RAW);
1598 		data->permitted &= CAP_TO_MASK(CAP_NET_RAW);
1599 
1600 		if (capset(&hdr, data) < 0)
1601 			fprintf(stderr, "Setting capabilities failed\n");
1602 
1603 		if (capget(&hdr, data)  < 0) {
1604 			fprintf(stderr, "Error getting capabilities\n");
1605 		} else {
1606 			fprintf(stderr, "Capabilities EFF %x Caps INH %x Caps Per %x\n",
1607 				data[0].effective, data[0].inheritable, data[0].permitted);
1608 			fprintf(stderr, "Capabilities EFF %x Caps INH %x Caps Per %x\n",
1609 				data[1].effective, data[1].inheritable, data[1].permitted);
1610 		}
1611 	} else {
1612 		if (setrlimit(RLIMIT_MEMLOCK, &r)) {
1613 			fprintf(stderr, "ERROR: setrlimit(RLIMIT_MEMLOCK) \"%s\"\n",
1614 				strerror(errno));
1615 			exit(EXIT_FAILURE);
1616 		}
1617 
1618 		if (opt_num_xsks > 1)
1619 			load_xdp_program(argv, &obj);
1620 	}
1621 
1622 	/* Reserve memory for the umem. Use hugepages if unaligned chunk mode */
1623 	bufs = mmap(NULL, NUM_FRAMES * opt_xsk_frame_size,
1624 		    PROT_READ | PROT_WRITE,
1625 		    MAP_PRIVATE | MAP_ANONYMOUS | opt_mmap_flags, -1, 0);
1626 	if (bufs == MAP_FAILED) {
1627 		printf("ERROR: mmap failed\n");
1628 		exit(EXIT_FAILURE);
1629 	}
1630 
1631 	/* Create sockets... */
1632 	umem = xsk_configure_umem(bufs, NUM_FRAMES * opt_xsk_frame_size);
1633 	if (opt_bench == BENCH_RXDROP || opt_bench == BENCH_L2FWD) {
1634 		rx = true;
1635 		xsk_populate_fill_ring(umem);
1636 	}
1637 	if (opt_bench == BENCH_L2FWD || opt_bench == BENCH_TXONLY)
1638 		tx = true;
1639 	for (i = 0; i < opt_num_xsks; i++)
1640 		xsks[num_socks++] = xsk_configure_socket(umem, rx, tx);
1641 
1642 	for (i = 0; i < opt_num_xsks; i++)
1643 		apply_setsockopt(xsks[i]);
1644 
1645 	if (opt_bench == BENCH_TXONLY) {
1646 		gen_eth_hdr_data();
1647 
1648 		for (i = 0; i < NUM_FRAMES; i++)
1649 			gen_eth_frame(umem, i * opt_xsk_frame_size);
1650 	}
1651 
1652 	if (opt_num_xsks > 1 && opt_bench != BENCH_TXONLY)
1653 		enter_xsks_into_map(obj);
1654 
1655 	if (opt_reduced_cap) {
1656 		ret = recv_xsks_map_fd(&xsks_map_fd);
1657 		if (ret) {
1658 			fprintf(stderr, "Error %d receiving xsks_map_fd\n", ret);
1659 			exit_with_error(ret);
1660 		}
1661 		if (xsks[0]->xsk) {
1662 			ret = xsk_socket__update_xskmap(xsks[0]->xsk, xsks_map_fd);
1663 			if (ret) {
1664 				fprintf(stderr, "Update of BPF map failed(%d)\n", ret);
1665 				exit_with_error(ret);
1666 			}
1667 		}
1668 	}
1669 
1670 	signal(SIGINT, int_exit);
1671 	signal(SIGTERM, int_exit);
1672 	signal(SIGABRT, int_exit);
1673 
1674 	setlocale(LC_ALL, "");
1675 
1676 	prev_time = get_nsecs();
1677 	start_time = prev_time;
1678 
1679 	if (!opt_quiet) {
1680 		ret = pthread_create(&pt, NULL, poller, NULL);
1681 		if (ret)
1682 			exit_with_error(ret);
1683 	}
1684 
1685 
1686 	if (opt_bench == BENCH_RXDROP)
1687 		rx_drop_all();
1688 	else if (opt_bench == BENCH_TXONLY)
1689 		tx_only_all();
1690 	else
1691 		l2fwd_all();
1692 
1693 	benchmark_done = true;
1694 
1695 	if (!opt_quiet)
1696 		pthread_join(pt, NULL);
1697 
1698 	xdpsock_cleanup();
1699 
1700 	munmap(bufs, NUM_FRAMES * opt_xsk_frame_size);
1701 
1702 	return 0;
1703 }
1704