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/limits.h>
15 #include <linux/udp.h>
16 #include <arpa/inet.h>
17 #include <locale.h>
18 #include <net/ethernet.h>
19 #include <net/if.h>
20 #include <poll.h>
21 #include <pthread.h>
22 #include <signal.h>
23 #include <stdbool.h>
24 #include <stdio.h>
25 #include <stdlib.h>
26 #include <string.h>
27 #include <sys/capability.h>
28 #include <sys/mman.h>
29 #include <sys/resource.h>
30 #include <sys/socket.h>
31 #include <sys/types.h>
32 #include <sys/un.h>
33 #include <time.h>
34 #include <unistd.h>
35
36 #include <bpf/libbpf.h>
37 #include <bpf/xsk.h>
38 #include <bpf/bpf.h>
39 #include "xdpsock.h"
40
41 #ifndef SOL_XDP
42 #define SOL_XDP 283
43 #endif
44
45 #ifndef AF_XDP
46 #define AF_XDP 44
47 #endif
48
49 #ifndef PF_XDP
50 #define PF_XDP AF_XDP
51 #endif
52
53 #define NUM_FRAMES (4 * 1024)
54 #define MIN_PKT_SIZE 64
55
56 #define DEBUG_HEXDUMP 0
57
58 typedef __u64 u64;
59 typedef __u32 u32;
60 typedef __u16 u16;
61 typedef __u8 u8;
62
63 static unsigned long prev_time;
64
65 enum benchmark_type {
66 BENCH_RXDROP = 0,
67 BENCH_TXONLY = 1,
68 BENCH_L2FWD = 2,
69 };
70
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;
75 static int opt_queue;
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 = "";
87 static u32 irq_no;
88 static int irqs_at_init = -1;
89 static int opt_poll;
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;
99 static u32 prog_id;
100 static bool opt_busy_poll;
101 static bool opt_reduced_cap;
102
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;
120 };
121
122 struct xsk_driver_stats {
123 unsigned long intrs;
124 unsigned long prev_intrs;
125 };
126
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;
138 };
139
140 struct xsk_umem_info {
141 struct xsk_ring_prod fq;
142 struct xsk_ring_cons cq;
143 struct xsk_umem *umem;
144 void *buffer;
145 };
146
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;
155 u32 outstanding_tx;
156 };
157
158 static int num_socks;
159 struct xsk_socket_info *xsks[MAX_SOCKS];
160 int sock;
161
162 static unsigned long get_nsecs(void)
163 {
164 struct timespec ts;
165
166 clock_gettime(CLOCK_MONOTONIC, &ts);
167 return ts.tv_sec * 1000000000UL + ts.tv_nsec;
168 }
169
170 static void print_benchmark(bool running)
171 {
172 const char *bench_str = "INVALID";
173
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)
179 bench_str = "l2fwd";
180
181 printf("%s:%d %s ", opt_if, opt_queue, bench_str);
182 if (opt_xdp_flags & XDP_FLAGS_SKB_MODE)
183 printf("xdp-skb ");
184 else if (opt_xdp_flags & XDP_FLAGS_DRV_MODE)
185 printf("xdp-drv ");
186 else
187 printf(" ");
188
189 if (opt_poll)
190 printf("poll() ");
191
192 if (running) {
193 printf("running...");
194 fflush(stdout);
195 }
196 }
197
198 static int xsk_get_xdp_stats(int fd, struct xsk_socket_info *xsk)
199 {
200 struct xdp_statistics stats;
201 socklen_t optlen;
202 int err;
203
204 optlen = sizeof(stats);
205 err = getsockopt(fd, SOL_XDP, XDP_STATISTICS, &stats, &optlen);
206 if (err)
207 return err;
208
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;
216 return 0;
217 }
218
219 return -EINVAL;
220 }
221
222 static void dump_app_stats(long dt)
223 {
224 int i;
225
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;
230
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)
239 * 1000000000. / dt;
240 opt_polls_ps = (xsks[i]->app_stats.opt_polls -
241 xsks[i]->app_stats.prev_opt_polls) * 1000000000. / dt;
242
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);
252
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;
258 }
259 }
260
261 static bool get_interrupt_number(void)
262 {
263 FILE *f_int_proc;
264 char line[4096];
265 bool found = false;
266
267 f_int_proc = fopen("/proc/interrupts", "r");
268 if (f_int_proc == NULL) {
269 printf("Failed to open /proc/interrupts.\n");
270 return found;
271 }
272
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");
278 break;
279 }
280
281 /* Extract interrupt number from line */
282 if (strstr(line, opt_irq_str) != NULL) {
283 irq_no = atoi(line);
284 found = true;
285 break;
286 }
287 }
288
289 fclose(f_int_proc);
290
291 return found;
292 }
293
294 static int get_irqs(void)
295 {
296 char count_path[PATH_MAX];
297 int total_intrs = -1;
298 FILE *f_count_proc;
299 char line[4096];
300
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);
306 return total_intrs;
307 }
308
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);
312 } else {
313 static const char com[2] = ",";
314 char *token;
315
316 total_intrs = 0;
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);
322 }
323 }
324
325 fclose(f_count_proc);
326
327 return total_intrs;
328 }
329
330 static void dump_driver_stats(long dt)
331 {
332 int i;
333
334 for (i = 0; i < num_socks && xsks[i]; i++) {
335 char *fmt = "%-18s %'-14.0f %'-14lu\n";
336 double intrs_ps;
337 int n_ints = get_irqs();
338
339 if (n_ints < 0) {
340 printf("error getting intr info for intr %i\n", irq_no);
341 return;
342 }
343 xsks[i]->drv_stats.intrs = n_ints - irqs_at_init;
344
345 intrs_ps = (xsks[i]->drv_stats.intrs - xsks[i]->drv_stats.prev_intrs) *
346 1000000000. / dt;
347
348 printf("\n%-18s %-14s %-14s\n", "", "intrs/s", "count");
349 printf(fmt, "irqs", intrs_ps, xsks[i]->drv_stats.intrs);
350
351 xsks[i]->drv_stats.prev_intrs = xsks[i]->drv_stats.intrs;
352 }
353 }
354
355 static void dump_stats(void)
356 {
357 unsigned long now = get_nsecs();
358 long dt = now - prev_time;
359 int i;
360
361 prev_time = now;
362
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;
367
368 rx_pps = (xsks[i]->ring_stats.rx_npkts - xsks[i]->ring_stats.prev_rx_npkts) *
369 1000000000. / dt;
370 tx_pps = (xsks[i]->ring_stats.tx_npkts - xsks[i]->ring_stats.prev_tx_npkts) *
371 1000000000. / dt;
372
373 printf("\n sock%d@", i);
374 print_benchmark(false);
375 printf("\n");
376
377 printf("%-18s %-14s %-14s %-14.2f\n", "", "pps", "pkts",
378 dt / 1000000000.);
379 printf(fmt, "rx", rx_pps, xsks[i]->ring_stats.rx_npkts);
380 printf(fmt, "tx", tx_pps, xsks[i]->ring_stats.tx_npkts);
381
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;
384
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) *
389 1000000000. / dt;
390 rx_invalid_pps = (xsks[i]->ring_stats.rx_invalid_npkts -
391 xsks[i]->ring_stats.prev_rx_invalid_npkts) *
392 1000000000. / dt;
393 tx_invalid_pps = (xsks[i]->ring_stats.tx_invalid_npkts -
394 xsks[i]->ring_stats.prev_tx_invalid_npkts) *
395 1000000000. / dt;
396 full_pps = (xsks[i]->ring_stats.rx_full_npkts -
397 xsks[i]->ring_stats.prev_rx_full_npkts) *
398 1000000000. / dt;
399 fill_empty_pps = (xsks[i]->ring_stats.rx_fill_empty_npkts -
400 xsks[i]->ring_stats.prev_rx_fill_empty_npkts) *
401 1000000000. / dt;
402 tx_empty_pps = (xsks[i]->ring_stats.tx_empty_npkts -
403 xsks[i]->ring_stats.prev_tx_empty_npkts) *
404 1000000000. / dt;
405
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);
418
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;
431 } else {
432 printf("%-15s\n", "Error retrieving extra stats");
433 }
434 }
435 }
436
437 if (opt_app_stats)
438 dump_app_stats(dt);
439 if (irq_no)
440 dump_driver_stats(dt);
441 }
442
443 static bool is_benchmark_done(void)
444 {
445 if (opt_duration > 0) {
446 unsigned long dt = (get_nsecs() - start_time);
447
448 if (dt >= opt_duration)
449 benchmark_done = true;
450 }
451 return benchmark_done;
452 }
453
454 static void *poller(void *arg)
455 {
456 (void)arg;
457 while (!is_benchmark_done()) {
458 sleep(opt_interval);
459 dump_stats();
460 }
461
462 return NULL;
463 }
464
465 static void remove_xdp_program(void)
466 {
467 u32 curr_prog_id = 0;
468 int cmd = CLOSE_CONN;
469
470 if (bpf_get_link_xdp_id(opt_ifindex, &curr_prog_id, opt_xdp_flags)) {
471 printf("bpf_get_link_xdp_id failed\n");
472 exit(EXIT_FAILURE);
473 }
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");
478 else
479 printf("program on interface changed, not removing\n");
480
481 if (opt_reduced_cap) {
482 if (write(sock, &cmd, sizeof(int)) < 0) {
483 fprintf(stderr, "Error writing into stream socket: %s", strerror(errno));
484 exit(EXIT_FAILURE);
485 }
486 }
487 }
488
489 static void int_exit(int sig)
490 {
491 benchmark_done = true;
492 }
493
494 static void xdpsock_cleanup(void)
495 {
496 struct xsk_umem *umem = xsks[0]->umem->umem;
497 int i;
498
499 dump_stats();
500 for (i = 0; i < num_socks; i++)
501 xsk_socket__delete(xsks[i]->xsk);
502 (void)xsk_umem__delete(umem);
503 remove_xdp_program();
504 }
505
506 static void __exit_with_error(int error, const char *file, const char *func,
507 int line)
508 {
509 fprintf(stderr, "%s:%s:%i: errno: %d/\"%s\"\n", file, func,
510 line, error, strerror(error));
511 dump_stats();
512 remove_xdp_program();
513 exit(EXIT_FAILURE);
514 }
515
516 #define exit_with_error(error) __exit_with_error(error, __FILE__, __func__, \
517 __LINE__)
518 static void swap_mac_addresses(void *data)
519 {
520 struct ether_header *eth = (struct ether_header *)data;
521 struct ether_addr *src_addr = (struct ether_addr *)ð->ether_shost;
522 struct ether_addr *dst_addr = (struct ether_addr *)ð->ether_dhost;
523 struct ether_addr tmp;
524
525 tmp = *src_addr;
526 *src_addr = *dst_addr;
527 *dst_addr = tmp;
528 }
529
530 static void hex_dump(void *pkt, size_t length, u64 addr)
531 {
532 const unsigned char *address = (unsigned char *)pkt;
533 const unsigned char *line = address;
534 size_t line_size = 32;
535 unsigned char c;
536 char buf[32];
537 int i = 0;
538
539 if (!DEBUG_HEXDUMP)
540 return;
541
542 sprintf(buf, "addr=%llu", addr);
543 printf("length = %zu\n", length);
544 printf("%s | ", buf);
545 while (length-- > 0) {
546 printf("%02X ", *address++);
547 if (!(++i % line_size) || (length == 0 && i % line_size)) {
548 if (length == 0) {
549 while (i++ % line_size)
550 printf("__ ");
551 }
552 printf(" | "); /* right close */
553 while (line < address) {
554 c = *line++;
555 printf("%c", (c < 33 || c == 255) ? 0x2E : c);
556 }
557 printf("\n");
558 if (length > 0)
559 printf("%s | ", buf);
560 }
561 }
562 printf("\n");
563 }
564
565 static void *memset32_htonl(void *dest, u32 val, u32 size)
566 {
567 u32 *ptr = (u32 *)dest;
568 int i;
569
570 val = htonl(val);
571
572 for (i = 0; i < (size & (~0x3)); i += 4)
573 ptr[i >> 2] = val;
574
575 for (; i < size; i++)
576 ((char *)dest)[i] = ((char *)&val)[i & 3];
577
578 return dest;
579 }
580
581 /*
582 * This function code has been taken from
583 * Linux kernel lib/checksum.c
584 */
585 static inline unsigned short from32to16(unsigned int x)
586 {
587 /* add up 16-bit and 16-bit for 16+c bit */
588 x = (x & 0xffff) + (x >> 16);
589 /* add up carry.. */
590 x = (x & 0xffff) + (x >> 16);
591 return x;
592 }
593
594 /*
595 * This function code has been taken from
596 * Linux kernel lib/checksum.c
597 */
598 static unsigned int do_csum(const unsigned char *buff, int len)
599 {
600 unsigned int result = 0;
601 int odd;
602
603 if (len <= 0)
604 goto out;
605 odd = 1 & (unsigned long)buff;
606 if (odd) {
607 #ifdef __LITTLE_ENDIAN
608 result += (*buff << 8);
609 #else
610 result = *buff;
611 #endif
612 len--;
613 buff++;
614 }
615 if (len >= 2) {
616 if (2 & (unsigned long)buff) {
617 result += *(unsigned short *)buff;
618 len -= 2;
619 buff += 2;
620 }
621 if (len >= 4) {
622 const unsigned char *end = buff +
623 ((unsigned int)len & ~3);
624 unsigned int carry = 0;
625
626 do {
627 unsigned int w = *(unsigned int *)buff;
628
629 buff += 4;
630 result += carry;
631 result += w;
632 carry = (w > result);
633 } while (buff < end);
634 result += carry;
635 result = (result & 0xffff) + (result >> 16);
636 }
637 if (len & 2) {
638 result += *(unsigned short *)buff;
639 buff += 2;
640 }
641 }
642 if (len & 1)
643 #ifdef __LITTLE_ENDIAN
644 result += *buff;
645 #else
646 result += (*buff << 8);
647 #endif
648 result = from32to16(result);
649 if (odd)
650 result = ((result >> 8) & 0xff) | ((result & 0xff) << 8);
651 out:
652 return result;
653 }
654
655 __sum16 ip_fast_csum(const void *iph, unsigned int ihl);
656
657 /*
658 * This is a version of ip_compute_csum() optimized for IP headers,
659 * which always checksum on 4 octet boundaries.
660 * This function code has been taken from
661 * Linux kernel lib/checksum.c
662 */
663 __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
664 {
665 return (__force __sum16)~do_csum(iph, ihl * 4);
666 }
667
668 /*
669 * Fold a partial checksum
670 * This function code has been taken from
671 * Linux kernel include/asm-generic/checksum.h
672 */
673 static inline __sum16 csum_fold(__wsum csum)
674 {
675 u32 sum = (__force u32)csum;
676
677 sum = (sum & 0xffff) + (sum >> 16);
678 sum = (sum & 0xffff) + (sum >> 16);
679 return (__force __sum16)~sum;
680 }
681
682 /*
683 * This function code has been taken from
684 * Linux kernel lib/checksum.c
685 */
686 static inline u32 from64to32(u64 x)
687 {
688 /* add up 32-bit and 32-bit for 32+c bit */
689 x = (x & 0xffffffff) + (x >> 32);
690 /* add up carry.. */
691 x = (x & 0xffffffff) + (x >> 32);
692 return (u32)x;
693 }
694
695 __wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
696 __u32 len, __u8 proto, __wsum sum);
697
698 /*
699 * This function code has been taken from
700 * Linux kernel lib/checksum.c
701 */
702 __wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
703 __u32 len, __u8 proto, __wsum sum)
704 {
705 unsigned long long s = (__force u32)sum;
706
707 s += (__force u32)saddr;
708 s += (__force u32)daddr;
709 #ifdef __BIG_ENDIAN__
710 s += proto + len;
711 #else
712 s += (proto + len) << 8;
713 #endif
714 return (__force __wsum)from64to32(s);
715 }
716
717 /*
718 * This function has been taken from
719 * Linux kernel include/asm-generic/checksum.h
720 */
721 static inline __sum16
722 csum_tcpudp_magic(__be32 saddr, __be32 daddr, __u32 len,
723 __u8 proto, __wsum sum)
724 {
725 return csum_fold(csum_tcpudp_nofold(saddr, daddr, len, proto, sum));
726 }
727
728 static inline u16 udp_csum(u32 saddr, u32 daddr, u32 len,
729 u8 proto, u16 *udp_pkt)
730 {
731 u32 csum = 0;
732 u32 cnt = 0;
733
734 /* udp hdr and data */
735 for (; cnt < len; cnt += 2)
736 csum += udp_pkt[cnt >> 1];
737
738 return csum_tcpudp_magic(saddr, daddr, len, proto, csum);
739 }
740
741 #define ETH_FCS_SIZE 4
742
743 #define PKT_HDR_SIZE (sizeof(struct ethhdr) + sizeof(struct iphdr) + \
744 sizeof(struct udphdr))
745
746 #define PKT_SIZE (opt_pkt_size - ETH_FCS_SIZE)
747 #define IP_PKT_SIZE (PKT_SIZE - sizeof(struct ethhdr))
748 #define UDP_PKT_SIZE (IP_PKT_SIZE - sizeof(struct iphdr))
749 #define UDP_PKT_DATA_SIZE (UDP_PKT_SIZE - sizeof(struct udphdr))
750
751 static u8 pkt_data[XSK_UMEM__DEFAULT_FRAME_SIZE];
752
753 static void gen_eth_hdr_data(void)
754 {
755 struct udphdr *udp_hdr = (struct udphdr *)(pkt_data +
756 sizeof(struct ethhdr) +
757 sizeof(struct iphdr));
758 struct iphdr *ip_hdr = (struct iphdr *)(pkt_data +
759 sizeof(struct ethhdr));
760 struct ethhdr *eth_hdr = (struct ethhdr *)pkt_data;
761
762 /* ethernet header */
763 memcpy(eth_hdr->h_dest, "\x3c\xfd\xfe\x9e\x7f\x71", ETH_ALEN);
764 memcpy(eth_hdr->h_source, "\xec\xb1\xd7\x98\x3a\xc0", ETH_ALEN);
765 eth_hdr->h_proto = htons(ETH_P_IP);
766
767 /* IP header */
768 ip_hdr->version = IPVERSION;
769 ip_hdr->ihl = 0x5; /* 20 byte header */
770 ip_hdr->tos = 0x0;
771 ip_hdr->tot_len = htons(IP_PKT_SIZE);
772 ip_hdr->id = 0;
773 ip_hdr->frag_off = 0;
774 ip_hdr->ttl = IPDEFTTL;
775 ip_hdr->protocol = IPPROTO_UDP;
776 ip_hdr->saddr = htonl(0x0a0a0a10);
777 ip_hdr->daddr = htonl(0x0a0a0a20);
778
779 /* IP header checksum */
780 ip_hdr->check = 0;
781 ip_hdr->check = ip_fast_csum((const void *)ip_hdr, ip_hdr->ihl);
782
783 /* UDP header */
784 udp_hdr->source = htons(0x1000);
785 udp_hdr->dest = htons(0x1000);
786 udp_hdr->len = htons(UDP_PKT_SIZE);
787
788 /* UDP data */
789 memset32_htonl(pkt_data + PKT_HDR_SIZE, opt_pkt_fill_pattern,
790 UDP_PKT_DATA_SIZE);
791
792 /* UDP header checksum */
793 udp_hdr->check = 0;
794 udp_hdr->check = udp_csum(ip_hdr->saddr, ip_hdr->daddr, UDP_PKT_SIZE,
795 IPPROTO_UDP, (u16 *)udp_hdr);
796 }
797
798 static void gen_eth_frame(struct xsk_umem_info *umem, u64 addr)
799 {
800 memcpy(xsk_umem__get_data(umem->buffer, addr), pkt_data,
801 PKT_SIZE);
802 }
803
804 static struct xsk_umem_info *xsk_configure_umem(void *buffer, u64 size)
805 {
806 struct xsk_umem_info *umem;
807 struct xsk_umem_config cfg = {
808 /* We recommend that you set the fill ring size >= HW RX ring size +
809 * AF_XDP RX ring size. Make sure you fill up the fill ring
810 * with buffers at regular intervals, and you will with this setting
811 * avoid allocation failures in the driver. These are usually quite
812 * expensive since drivers have not been written to assume that
813 * allocation failures are common. For regular sockets, kernel
814 * allocated memory is used that only runs out in OOM situations
815 * that should be rare.
816 */
817 .fill_size = XSK_RING_PROD__DEFAULT_NUM_DESCS * 2,
818 .comp_size = XSK_RING_CONS__DEFAULT_NUM_DESCS,
819 .frame_size = opt_xsk_frame_size,
820 .frame_headroom = XSK_UMEM__DEFAULT_FRAME_HEADROOM,
821 .flags = opt_umem_flags
822 };
823 int ret;
824
825 umem = calloc(1, sizeof(*umem));
826 if (!umem)
827 exit_with_error(errno);
828
829 ret = xsk_umem__create(&umem->umem, buffer, size, &umem->fq, &umem->cq,
830 &cfg);
831 if (ret)
832 exit_with_error(-ret);
833
834 umem->buffer = buffer;
835 return umem;
836 }
837
838 static void xsk_populate_fill_ring(struct xsk_umem_info *umem)
839 {
840 int ret, i;
841 u32 idx;
842
843 ret = xsk_ring_prod__reserve(&umem->fq,
844 XSK_RING_PROD__DEFAULT_NUM_DESCS * 2, &idx);
845 if (ret != XSK_RING_PROD__DEFAULT_NUM_DESCS * 2)
846 exit_with_error(-ret);
847 for (i = 0; i < XSK_RING_PROD__DEFAULT_NUM_DESCS * 2; i++)
848 *xsk_ring_prod__fill_addr(&umem->fq, idx++) =
849 i * opt_xsk_frame_size;
850 xsk_ring_prod__submit(&umem->fq, XSK_RING_PROD__DEFAULT_NUM_DESCS * 2);
851 }
852
853 static struct xsk_socket_info *xsk_configure_socket(struct xsk_umem_info *umem,
854 bool rx, bool tx)
855 {
856 struct xsk_socket_config cfg;
857 struct xsk_socket_info *xsk;
858 struct xsk_ring_cons *rxr;
859 struct xsk_ring_prod *txr;
860 int ret;
861
862 xsk = calloc(1, sizeof(*xsk));
863 if (!xsk)
864 exit_with_error(errno);
865
866 xsk->umem = umem;
867 cfg.rx_size = XSK_RING_CONS__DEFAULT_NUM_DESCS;
868 cfg.tx_size = XSK_RING_PROD__DEFAULT_NUM_DESCS;
869 if (opt_num_xsks > 1 || opt_reduced_cap)
870 cfg.libbpf_flags = XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD;
871 else
872 cfg.libbpf_flags = 0;
873 cfg.xdp_flags = opt_xdp_flags;
874 cfg.bind_flags = opt_xdp_bind_flags;
875
876 rxr = rx ? &xsk->rx : NULL;
877 txr = tx ? &xsk->tx : NULL;
878 ret = xsk_socket__create(&xsk->xsk, opt_if, opt_queue, umem->umem,
879 rxr, txr, &cfg);
880 if (ret)
881 exit_with_error(-ret);
882
883 ret = bpf_get_link_xdp_id(opt_ifindex, &prog_id, opt_xdp_flags);
884 if (ret)
885 exit_with_error(-ret);
886
887 xsk->app_stats.rx_empty_polls = 0;
888 xsk->app_stats.fill_fail_polls = 0;
889 xsk->app_stats.copy_tx_sendtos = 0;
890 xsk->app_stats.tx_wakeup_sendtos = 0;
891 xsk->app_stats.opt_polls = 0;
892 xsk->app_stats.prev_rx_empty_polls = 0;
893 xsk->app_stats.prev_fill_fail_polls = 0;
894 xsk->app_stats.prev_copy_tx_sendtos = 0;
895 xsk->app_stats.prev_tx_wakeup_sendtos = 0;
896 xsk->app_stats.prev_opt_polls = 0;
897
898 return xsk;
899 }
900
901 static struct option long_options[] = {
902 {"rxdrop", no_argument, 0, 'r'},
903 {"txonly", no_argument, 0, 't'},
904 {"l2fwd", no_argument, 0, 'l'},
905 {"interface", required_argument, 0, 'i'},
906 {"queue", required_argument, 0, 'q'},
907 {"poll", no_argument, 0, 'p'},
908 {"xdp-skb", no_argument, 0, 'S'},
909 {"xdp-native", no_argument, 0, 'N'},
910 {"interval", required_argument, 0, 'n'},
911 {"zero-copy", no_argument, 0, 'z'},
912 {"copy", no_argument, 0, 'c'},
913 {"frame-size", required_argument, 0, 'f'},
914 {"no-need-wakeup", no_argument, 0, 'm'},
915 {"unaligned", no_argument, 0, 'u'},
916 {"shared-umem", no_argument, 0, 'M'},
917 {"force", no_argument, 0, 'F'},
918 {"duration", required_argument, 0, 'd'},
919 {"batch-size", required_argument, 0, 'b'},
920 {"tx-pkt-count", required_argument, 0, 'C'},
921 {"tx-pkt-size", required_argument, 0, 's'},
922 {"tx-pkt-pattern", required_argument, 0, 'P'},
923 {"extra-stats", no_argument, 0, 'x'},
924 {"quiet", no_argument, 0, 'Q'},
925 {"app-stats", no_argument, 0, 'a'},
926 {"irq-string", no_argument, 0, 'I'},
927 {"busy-poll", no_argument, 0, 'B'},
928 {"reduce-cap", no_argument, 0, 'R'},
929 {0, 0, 0, 0}
930 };
931
932 static void usage(const char *prog)
933 {
934 const char *str =
935 " Usage: %s [OPTIONS]\n"
936 " Options:\n"
937 " -r, --rxdrop Discard all incoming packets (default)\n"
938 " -t, --txonly Only send packets\n"
939 " -l, --l2fwd MAC swap L2 forwarding\n"
940 " -i, --interface=n Run on interface n\n"
941 " -q, --queue=n Use queue n (default 0)\n"
942 " -p, --poll Use poll syscall\n"
943 " -S, --xdp-skb=n Use XDP skb-mod\n"
944 " -N, --xdp-native=n Enforce XDP native mode\n"
945 " -n, --interval=n Specify statistics update interval (default 1 sec).\n"
946 " -z, --zero-copy Force zero-copy mode.\n"
947 " -c, --copy Force copy mode.\n"
948 " -m, --no-need-wakeup Turn off use of driver need wakeup flag.\n"
949 " -f, --frame-size=n Set the frame size (must be a power of two in aligned mode, default is %d).\n"
950 " -u, --unaligned Enable unaligned chunk placement\n"
951 " -M, --shared-umem Enable XDP_SHARED_UMEM (cannot be used with -R)\n"
952 " -F, --force Force loading the XDP prog\n"
953 " -d, --duration=n Duration in secs to run command.\n"
954 " Default: forever.\n"
955 " -b, --batch-size=n Batch size for sending or receiving\n"
956 " packets. Default: %d\n"
957 " -C, --tx-pkt-count=n Number of packets to send.\n"
958 " Default: Continuous packets.\n"
959 " -s, --tx-pkt-size=n Transmit packet size.\n"
960 " (Default: %d bytes)\n"
961 " Min size: %d, Max size %d.\n"
962 " -P, --tx-pkt-pattern=nPacket fill pattern. Default: 0x%x\n"
963 " -x, --extra-stats Display extra statistics.\n"
964 " -Q, --quiet Do not display any stats.\n"
965 " -a, --app-stats Display application (syscall) statistics.\n"
966 " -I, --irq-string Display driver interrupt statistics for interface associated with irq-string.\n"
967 " -B, --busy-poll Busy poll.\n"
968 " -R, --reduce-cap Use reduced capabilities (cannot be used with -M)\n"
969 "\n";
970 fprintf(stderr, str, prog, XSK_UMEM__DEFAULT_FRAME_SIZE,
971 opt_batch_size, MIN_PKT_SIZE, MIN_PKT_SIZE,
972 XSK_UMEM__DEFAULT_FRAME_SIZE, opt_pkt_fill_pattern);
973
974 exit(EXIT_FAILURE);
975 }
976
977 static void parse_command_line(int argc, char **argv)
978 {
979 int option_index, c;
980
981 opterr = 0;
982
983 for (;;) {
984 c = getopt_long(argc, argv, "Frtli:q:pSNn:czf:muMd:b:C:s:P:xQaI:BR",
985 long_options, &option_index);
986 if (c == -1)
987 break;
988
989 switch (c) {
990 case 'r':
991 opt_bench = BENCH_RXDROP;
992 break;
993 case 't':
994 opt_bench = BENCH_TXONLY;
995 break;
996 case 'l':
997 opt_bench = BENCH_L2FWD;
998 break;
999 case 'i':
1000 opt_if = optarg;
1001 break;
1002 case 'q':
1003 opt_queue = atoi(optarg);
1004 break;
1005 case 'p':
1006 opt_poll = 1;
1007 break;
1008 case 'S':
1009 opt_xdp_flags |= XDP_FLAGS_SKB_MODE;
1010 opt_xdp_bind_flags |= XDP_COPY;
1011 break;
1012 case 'N':
1013 /* default, set below */
1014 break;
1015 case 'n':
1016 opt_interval = atoi(optarg);
1017 break;
1018 case 'z':
1019 opt_xdp_bind_flags |= XDP_ZEROCOPY;
1020 break;
1021 case 'c':
1022 opt_xdp_bind_flags |= XDP_COPY;
1023 break;
1024 case 'u':
1025 opt_umem_flags |= XDP_UMEM_UNALIGNED_CHUNK_FLAG;
1026 opt_unaligned_chunks = 1;
1027 opt_mmap_flags = MAP_HUGETLB;
1028 break;
1029 case 'F':
1030 opt_xdp_flags &= ~XDP_FLAGS_UPDATE_IF_NOEXIST;
1031 break;
1032 case 'f':
1033 opt_xsk_frame_size = atoi(optarg);
1034 break;
1035 case 'm':
1036 opt_need_wakeup = false;
1037 opt_xdp_bind_flags &= ~XDP_USE_NEED_WAKEUP;
1038 break;
1039 case 'M':
1040 opt_num_xsks = MAX_SOCKS;
1041 break;
1042 case 'd':
1043 opt_duration = atoi(optarg);
1044 opt_duration *= 1000000000;
1045 break;
1046 case 'b':
1047 opt_batch_size = atoi(optarg);
1048 break;
1049 case 'C':
1050 opt_pkt_count = atoi(optarg);
1051 break;
1052 case 's':
1053 opt_pkt_size = atoi(optarg);
1054 if (opt_pkt_size > (XSK_UMEM__DEFAULT_FRAME_SIZE) ||
1055 opt_pkt_size < MIN_PKT_SIZE) {
1056 fprintf(stderr,
1057 "ERROR: Invalid frame size %d\n",
1058 opt_pkt_size);
1059 usage(basename(argv[0]));
1060 }
1061 break;
1062 case 'P':
1063 opt_pkt_fill_pattern = strtol(optarg, NULL, 16);
1064 break;
1065 case 'x':
1066 opt_extra_stats = 1;
1067 break;
1068 case 'Q':
1069 opt_quiet = 1;
1070 break;
1071 case 'a':
1072 opt_app_stats = 1;
1073 break;
1074 case 'I':
1075 opt_irq_str = optarg;
1076 if (get_interrupt_number())
1077 irqs_at_init = get_irqs();
1078 if (irqs_at_init < 0) {
1079 fprintf(stderr, "ERROR: Failed to get irqs for %s\n", opt_irq_str);
1080 usage(basename(argv[0]));
1081 }
1082 break;
1083 case 'B':
1084 opt_busy_poll = 1;
1085 break;
1086 case 'R':
1087 opt_reduced_cap = true;
1088 break;
1089 default:
1090 usage(basename(argv[0]));
1091 }
1092 }
1093
1094 if (!(opt_xdp_flags & XDP_FLAGS_SKB_MODE))
1095 opt_xdp_flags |= XDP_FLAGS_DRV_MODE;
1096
1097 opt_ifindex = if_nametoindex(opt_if);
1098 if (!opt_ifindex) {
1099 fprintf(stderr, "ERROR: interface \"%s\" does not exist\n",
1100 opt_if);
1101 usage(basename(argv[0]));
1102 }
1103
1104 if ((opt_xsk_frame_size & (opt_xsk_frame_size - 1)) &&
1105 !opt_unaligned_chunks) {
1106 fprintf(stderr, "--frame-size=%d is not a power of two\n",
1107 opt_xsk_frame_size);
1108 usage(basename(argv[0]));
1109 }
1110
1111 if (opt_reduced_cap && opt_num_xsks > 1) {
1112 fprintf(stderr, "ERROR: -M and -R cannot be used together\n");
1113 usage(basename(argv[0]));
1114 }
1115 }
1116
1117 static void kick_tx(struct xsk_socket_info *xsk)
1118 {
1119 int ret;
1120
1121 ret = sendto(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, 0);
1122 if (ret >= 0 || errno == ENOBUFS || errno == EAGAIN ||
1123 errno == EBUSY || errno == ENETDOWN)
1124 return;
1125 exit_with_error(errno);
1126 }
1127
1128 static inline void complete_tx_l2fwd(struct xsk_socket_info *xsk)
1129 {
1130 struct xsk_umem_info *umem = xsk->umem;
1131 u32 idx_cq = 0, idx_fq = 0;
1132 unsigned int rcvd;
1133 size_t ndescs;
1134
1135 if (!xsk->outstanding_tx)
1136 return;
1137
1138 /* In copy mode, Tx is driven by a syscall so we need to use e.g. sendto() to
1139 * really send the packets. In zero-copy mode we do not have to do this, since Tx
1140 * is driven by the NAPI loop. So as an optimization, we do not have to call
1141 * sendto() all the time in zero-copy mode for l2fwd.
1142 */
1143 if (opt_xdp_bind_flags & XDP_COPY) {
1144 xsk->app_stats.copy_tx_sendtos++;
1145 kick_tx(xsk);
1146 }
1147
1148 ndescs = (xsk->outstanding_tx > opt_batch_size) ? opt_batch_size :
1149 xsk->outstanding_tx;
1150
1151 /* re-add completed Tx buffers */
1152 rcvd = xsk_ring_cons__peek(&umem->cq, ndescs, &idx_cq);
1153 if (rcvd > 0) {
1154 unsigned int i;
1155 int ret;
1156
1157 ret = xsk_ring_prod__reserve(&umem->fq, rcvd, &idx_fq);
1158 while (ret != rcvd) {
1159 if (ret < 0)
1160 exit_with_error(-ret);
1161 if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&umem->fq)) {
1162 xsk->app_stats.fill_fail_polls++;
1163 recvfrom(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL,
1164 NULL);
1165 }
1166 ret = xsk_ring_prod__reserve(&umem->fq, rcvd, &idx_fq);
1167 }
1168
1169 for (i = 0; i < rcvd; i++)
1170 *xsk_ring_prod__fill_addr(&umem->fq, idx_fq++) =
1171 *xsk_ring_cons__comp_addr(&umem->cq, idx_cq++);
1172
1173 xsk_ring_prod__submit(&xsk->umem->fq, rcvd);
1174 xsk_ring_cons__release(&xsk->umem->cq, rcvd);
1175 xsk->outstanding_tx -= rcvd;
1176 }
1177 }
1178
1179 static inline void complete_tx_only(struct xsk_socket_info *xsk,
1180 int batch_size)
1181 {
1182 unsigned int rcvd;
1183 u32 idx;
1184
1185 if (!xsk->outstanding_tx)
1186 return;
1187
1188 if (!opt_need_wakeup || xsk_ring_prod__needs_wakeup(&xsk->tx)) {
1189 xsk->app_stats.tx_wakeup_sendtos++;
1190 kick_tx(xsk);
1191 }
1192
1193 rcvd = xsk_ring_cons__peek(&xsk->umem->cq, batch_size, &idx);
1194 if (rcvd > 0) {
1195 xsk_ring_cons__release(&xsk->umem->cq, rcvd);
1196 xsk->outstanding_tx -= rcvd;
1197 }
1198 }
1199
1200 static void rx_drop(struct xsk_socket_info *xsk)
1201 {
1202 unsigned int rcvd, i;
1203 u32 idx_rx = 0, idx_fq = 0;
1204 int ret;
1205
1206 rcvd = xsk_ring_cons__peek(&xsk->rx, opt_batch_size, &idx_rx);
1207 if (!rcvd) {
1208 if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&xsk->umem->fq)) {
1209 xsk->app_stats.rx_empty_polls++;
1210 recvfrom(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, NULL);
1211 }
1212 return;
1213 }
1214
1215 ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
1216 while (ret != rcvd) {
1217 if (ret < 0)
1218 exit_with_error(-ret);
1219 if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&xsk->umem->fq)) {
1220 xsk->app_stats.fill_fail_polls++;
1221 recvfrom(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, NULL);
1222 }
1223 ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
1224 }
1225
1226 for (i = 0; i < rcvd; i++) {
1227 u64 addr = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx)->addr;
1228 u32 len = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx++)->len;
1229 u64 orig = xsk_umem__extract_addr(addr);
1230
1231 addr = xsk_umem__add_offset_to_addr(addr);
1232 char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
1233
1234 hex_dump(pkt, len, addr);
1235 *xsk_ring_prod__fill_addr(&xsk->umem->fq, idx_fq++) = orig;
1236 }
1237
1238 xsk_ring_prod__submit(&xsk->umem->fq, rcvd);
1239 xsk_ring_cons__release(&xsk->rx, rcvd);
1240 xsk->ring_stats.rx_npkts += rcvd;
1241 }
1242
1243 static void rx_drop_all(void)
1244 {
1245 struct pollfd fds[MAX_SOCKS] = {};
1246 int i, ret;
1247
1248 for (i = 0; i < num_socks; i++) {
1249 fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
1250 fds[i].events = POLLIN;
1251 }
1252
1253 for (;;) {
1254 if (opt_poll) {
1255 for (i = 0; i < num_socks; i++)
1256 xsks[i]->app_stats.opt_polls++;
1257 ret = poll(fds, num_socks, opt_timeout);
1258 if (ret <= 0)
1259 continue;
1260 }
1261
1262 for (i = 0; i < num_socks; i++)
1263 rx_drop(xsks[i]);
1264
1265 if (benchmark_done)
1266 break;
1267 }
1268 }
1269
1270 static void tx_only(struct xsk_socket_info *xsk, u32 *frame_nb, int batch_size)
1271 {
1272 u32 idx;
1273 unsigned int i;
1274
1275 while (xsk_ring_prod__reserve(&xsk->tx, batch_size, &idx) <
1276 batch_size) {
1277 complete_tx_only(xsk, batch_size);
1278 if (benchmark_done)
1279 return;
1280 }
1281
1282 for (i = 0; i < batch_size; i++) {
1283 struct xdp_desc *tx_desc = xsk_ring_prod__tx_desc(&xsk->tx,
1284 idx + i);
1285 tx_desc->addr = (*frame_nb + i) << XSK_UMEM__DEFAULT_FRAME_SHIFT;
1286 tx_desc->len = PKT_SIZE;
1287 }
1288
1289 xsk_ring_prod__submit(&xsk->tx, batch_size);
1290 xsk->ring_stats.tx_npkts += batch_size;
1291 xsk->outstanding_tx += batch_size;
1292 *frame_nb += batch_size;
1293 *frame_nb %= NUM_FRAMES;
1294 complete_tx_only(xsk, batch_size);
1295 }
1296
1297 static inline int get_batch_size(int pkt_cnt)
1298 {
1299 if (!opt_pkt_count)
1300 return opt_batch_size;
1301
1302 if (pkt_cnt + opt_batch_size <= opt_pkt_count)
1303 return opt_batch_size;
1304
1305 return opt_pkt_count - pkt_cnt;
1306 }
1307
1308 static void complete_tx_only_all(void)
1309 {
1310 bool pending;
1311 int i;
1312
1313 do {
1314 pending = false;
1315 for (i = 0; i < num_socks; i++) {
1316 if (xsks[i]->outstanding_tx) {
1317 complete_tx_only(xsks[i], opt_batch_size);
1318 pending = !!xsks[i]->outstanding_tx;
1319 }
1320 }
1321 } while (pending);
1322 }
1323
1324 static void tx_only_all(void)
1325 {
1326 struct pollfd fds[MAX_SOCKS] = {};
1327 u32 frame_nb[MAX_SOCKS] = {};
1328 int pkt_cnt = 0;
1329 int i, ret;
1330
1331 for (i = 0; i < num_socks; i++) {
1332 fds[0].fd = xsk_socket__fd(xsks[i]->xsk);
1333 fds[0].events = POLLOUT;
1334 }
1335
1336 while ((opt_pkt_count && pkt_cnt < opt_pkt_count) || !opt_pkt_count) {
1337 int batch_size = get_batch_size(pkt_cnt);
1338
1339 if (opt_poll) {
1340 for (i = 0; i < num_socks; i++)
1341 xsks[i]->app_stats.opt_polls++;
1342 ret = poll(fds, num_socks, opt_timeout);
1343 if (ret <= 0)
1344 continue;
1345
1346 if (!(fds[0].revents & POLLOUT))
1347 continue;
1348 }
1349
1350 for (i = 0; i < num_socks; i++)
1351 tx_only(xsks[i], &frame_nb[i], batch_size);
1352
1353 pkt_cnt += batch_size;
1354
1355 if (benchmark_done)
1356 break;
1357 }
1358
1359 if (opt_pkt_count)
1360 complete_tx_only_all();
1361 }
1362
1363 static void l2fwd(struct xsk_socket_info *xsk)
1364 {
1365 unsigned int rcvd, i;
1366 u32 idx_rx = 0, idx_tx = 0;
1367 int ret;
1368
1369 complete_tx_l2fwd(xsk);
1370
1371 rcvd = xsk_ring_cons__peek(&xsk->rx, opt_batch_size, &idx_rx);
1372 if (!rcvd) {
1373 if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&xsk->umem->fq)) {
1374 xsk->app_stats.rx_empty_polls++;
1375 recvfrom(xsk_socket__fd(xsk->xsk), NULL, 0, MSG_DONTWAIT, NULL, NULL);
1376 }
1377 return;
1378 }
1379 xsk->ring_stats.rx_npkts += rcvd;
1380
1381 ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
1382 while (ret != rcvd) {
1383 if (ret < 0)
1384 exit_with_error(-ret);
1385 complete_tx_l2fwd(xsk);
1386 if (opt_busy_poll || xsk_ring_prod__needs_wakeup(&xsk->tx)) {
1387 xsk->app_stats.tx_wakeup_sendtos++;
1388 kick_tx(xsk);
1389 }
1390 ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx);
1391 }
1392
1393 for (i = 0; i < rcvd; i++) {
1394 u64 addr = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx)->addr;
1395 u32 len = xsk_ring_cons__rx_desc(&xsk->rx, idx_rx++)->len;
1396 u64 orig = addr;
1397
1398 addr = xsk_umem__add_offset_to_addr(addr);
1399 char *pkt = xsk_umem__get_data(xsk->umem->buffer, addr);
1400
1401 swap_mac_addresses(pkt);
1402
1403 hex_dump(pkt, len, addr);
1404 xsk_ring_prod__tx_desc(&xsk->tx, idx_tx)->addr = orig;
1405 xsk_ring_prod__tx_desc(&xsk->tx, idx_tx++)->len = len;
1406 }
1407
1408 xsk_ring_prod__submit(&xsk->tx, rcvd);
1409 xsk_ring_cons__release(&xsk->rx, rcvd);
1410
1411 xsk->ring_stats.tx_npkts += rcvd;
1412 xsk->outstanding_tx += rcvd;
1413 }
1414
1415 static void l2fwd_all(void)
1416 {
1417 struct pollfd fds[MAX_SOCKS] = {};
1418 int i, ret;
1419
1420 for (;;) {
1421 if (opt_poll) {
1422 for (i = 0; i < num_socks; i++) {
1423 fds[i].fd = xsk_socket__fd(xsks[i]->xsk);
1424 fds[i].events = POLLOUT | POLLIN;
1425 xsks[i]->app_stats.opt_polls++;
1426 }
1427 ret = poll(fds, num_socks, opt_timeout);
1428 if (ret <= 0)
1429 continue;
1430 }
1431
1432 for (i = 0; i < num_socks; i++)
1433 l2fwd(xsks[i]);
1434
1435 if (benchmark_done)
1436 break;
1437 }
1438 }
1439
1440 static void load_xdp_program(char **argv, struct bpf_object **obj)
1441 {
1442 struct bpf_prog_load_attr prog_load_attr = {
1443 .prog_type = BPF_PROG_TYPE_XDP,
1444 };
1445 char xdp_filename[256];
1446 int prog_fd;
1447
1448 snprintf(xdp_filename, sizeof(xdp_filename), "%s_kern.o", argv[0]);
1449 prog_load_attr.file = xdp_filename;
1450
1451 if (bpf_prog_load_xattr(&prog_load_attr, obj, &prog_fd))
1452 exit(EXIT_FAILURE);
1453 if (prog_fd < 0) {
1454 fprintf(stderr, "ERROR: no program found: %s\n",
1455 strerror(prog_fd));
1456 exit(EXIT_FAILURE);
1457 }
1458
1459 if (bpf_set_link_xdp_fd(opt_ifindex, prog_fd, opt_xdp_flags) < 0) {
1460 fprintf(stderr, "ERROR: link set xdp fd failed\n");
1461 exit(EXIT_FAILURE);
1462 }
1463 }
1464
1465 static void enter_xsks_into_map(struct bpf_object *obj)
1466 {
1467 struct bpf_map *map;
1468 int i, xsks_map;
1469
1470 map = bpf_object__find_map_by_name(obj, "xsks_map");
1471 xsks_map = bpf_map__fd(map);
1472 if (xsks_map < 0) {
1473 fprintf(stderr, "ERROR: no xsks map found: %s\n",
1474 strerror(xsks_map));
1475 exit(EXIT_FAILURE);
1476 }
1477
1478 for (i = 0; i < num_socks; i++) {
1479 int fd = xsk_socket__fd(xsks[i]->xsk);
1480 int key, ret;
1481
1482 key = i;
1483 ret = bpf_map_update_elem(xsks_map, &key, &fd, 0);
1484 if (ret) {
1485 fprintf(stderr, "ERROR: bpf_map_update_elem %d\n", i);
1486 exit(EXIT_FAILURE);
1487 }
1488 }
1489 }
1490
1491 static void apply_setsockopt(struct xsk_socket_info *xsk)
1492 {
1493 int sock_opt;
1494
1495 if (!opt_busy_poll)
1496 return;
1497
1498 sock_opt = 1;
1499 if (setsockopt(xsk_socket__fd(xsk->xsk), SOL_SOCKET, SO_PREFER_BUSY_POLL,
1500 (void *)&sock_opt, sizeof(sock_opt)) < 0)
1501 exit_with_error(errno);
1502
1503 sock_opt = 20;
1504 if (setsockopt(xsk_socket__fd(xsk->xsk), SOL_SOCKET, SO_BUSY_POLL,
1505 (void *)&sock_opt, sizeof(sock_opt)) < 0)
1506 exit_with_error(errno);
1507
1508 sock_opt = opt_batch_size;
1509 if (setsockopt(xsk_socket__fd(xsk->xsk), SOL_SOCKET, SO_BUSY_POLL_BUDGET,
1510 (void *)&sock_opt, sizeof(sock_opt)) < 0)
1511 exit_with_error(errno);
1512 }
1513
1514 static int recv_xsks_map_fd_from_ctrl_node(int sock, int *_fd)
1515 {
1516 char cms[CMSG_SPACE(sizeof(int))];
1517 struct cmsghdr *cmsg;
1518 struct msghdr msg;
1519 struct iovec iov;
1520 int value;
1521 int len;
1522
1523 iov.iov_base = &value;
1524 iov.iov_len = sizeof(int);
1525
1526 msg.msg_name = 0;
1527 msg.msg_namelen = 0;
1528 msg.msg_iov = &iov;
1529 msg.msg_iovlen = 1;
1530 msg.msg_flags = 0;
1531 msg.msg_control = (caddr_t)cms;
1532 msg.msg_controllen = sizeof(cms);
1533
1534 len = recvmsg(sock, &msg, 0);
1535
1536 if (len < 0) {
1537 fprintf(stderr, "Recvmsg failed length incorrect.\n");
1538 return -EINVAL;
1539 }
1540
1541 if (len == 0) {
1542 fprintf(stderr, "Recvmsg failed no data\n");
1543 return -EINVAL;
1544 }
1545
1546 cmsg = CMSG_FIRSTHDR(&msg);
1547 *_fd = *(int *)CMSG_DATA(cmsg);
1548
1549 return 0;
1550 }
1551
1552 static int
1553 recv_xsks_map_fd(int *xsks_map_fd)
1554 {
1555 struct sockaddr_un server;
1556 int err;
1557
1558 sock = socket(AF_UNIX, SOCK_STREAM, 0);
1559 if (sock < 0) {
1560 fprintf(stderr, "Error opening socket stream: %s", strerror(errno));
1561 return errno;
1562 }
1563
1564 server.sun_family = AF_UNIX;
1565 strcpy(server.sun_path, SOCKET_NAME);
1566
1567 if (connect(sock, (struct sockaddr *)&server, sizeof(struct sockaddr_un)) < 0) {
1568 close(sock);
1569 fprintf(stderr, "Error connecting stream socket: %s", strerror(errno));
1570 return errno;
1571 }
1572
1573 err = recv_xsks_map_fd_from_ctrl_node(sock, xsks_map_fd);
1574 if (err) {
1575 fprintf(stderr, "Error %d receiving fd\n", err);
1576 return err;
1577 }
1578 return 0;
1579 }
1580
1581 int main(int argc, char **argv)
1582 {
1583 struct __user_cap_header_struct hdr = { _LINUX_CAPABILITY_VERSION_3, 0 };
1584 struct __user_cap_data_struct data[2] = { { 0 } };
1585 struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
1586 bool rx = false, tx = false;
1587 struct xsk_umem_info *umem;
1588 struct bpf_object *obj;
1589 int xsks_map_fd = 0;
1590 pthread_t pt;
1591 int i, ret;
1592 void *bufs;
1593
1594 parse_command_line(argc, argv);
1595
1596 if (opt_reduced_cap) {
1597 if (capget(&hdr, data) < 0)
1598 fprintf(stderr, "Error getting capabilities\n");
1599
1600 data->effective &= CAP_TO_MASK(CAP_NET_RAW);
1601 data->permitted &= CAP_TO_MASK(CAP_NET_RAW);
1602
1603 if (capset(&hdr, data) < 0)
1604 fprintf(stderr, "Setting capabilities failed\n");
1605
1606 if (capget(&hdr, data) < 0) {
1607 fprintf(stderr, "Error getting capabilities\n");
1608 } else {
1609 fprintf(stderr, "Capabilities EFF %x Caps INH %x Caps Per %x\n",
1610 data[0].effective, data[0].inheritable, data[0].permitted);
1611 fprintf(stderr, "Capabilities EFF %x Caps INH %x Caps Per %x\n",
1612 data[1].effective, data[1].inheritable, data[1].permitted);
1613 }
1614 } else {
1615 if (setrlimit(RLIMIT_MEMLOCK, &r)) {
1616 fprintf(stderr, "ERROR: setrlimit(RLIMIT_MEMLOCK) \"%s\"\n",
1617 strerror(errno));
1618 exit(EXIT_FAILURE);
1619 }
1620
1621 if (opt_num_xsks > 1)
1622 load_xdp_program(argv, &obj);
1623 }
1624
1625 /* Reserve memory for the umem. Use hugepages if unaligned chunk mode */
1626 bufs = mmap(NULL, NUM_FRAMES * opt_xsk_frame_size,
1627 PROT_READ | PROT_WRITE,
1628 MAP_PRIVATE | MAP_ANONYMOUS | opt_mmap_flags, -1, 0);
1629 if (bufs == MAP_FAILED) {
1630 printf("ERROR: mmap failed\n");
1631 exit(EXIT_FAILURE);
1632 }
1633
1634 /* Create sockets... */
1635 umem = xsk_configure_umem(bufs, NUM_FRAMES * opt_xsk_frame_size);
1636 if (opt_bench == BENCH_RXDROP || opt_bench == BENCH_L2FWD) {
1637 rx = true;
1638 xsk_populate_fill_ring(umem);
1639 }
1640 if (opt_bench == BENCH_L2FWD || opt_bench == BENCH_TXONLY)
1641 tx = true;
1642 for (i = 0; i < opt_num_xsks; i++)
1643 xsks[num_socks++] = xsk_configure_socket(umem, rx, tx);
1644
1645 for (i = 0; i < opt_num_xsks; i++)
1646 apply_setsockopt(xsks[i]);
1647
1648 if (opt_bench == BENCH_TXONLY) {
1649 gen_eth_hdr_data();
1650
1651 for (i = 0; i < NUM_FRAMES; i++)
1652 gen_eth_frame(umem, i * opt_xsk_frame_size);
1653 }
1654
1655 if (opt_num_xsks > 1 && opt_bench != BENCH_TXONLY)
1656 enter_xsks_into_map(obj);
1657
1658 if (opt_reduced_cap) {
1659 ret = recv_xsks_map_fd(&xsks_map_fd);
1660 if (ret) {
1661 fprintf(stderr, "Error %d receiving xsks_map_fd\n", ret);
1662 exit_with_error(ret);
1663 }
1664 if (xsks[0]->xsk) {
1665 ret = xsk_socket__update_xskmap(xsks[0]->xsk, xsks_map_fd);
1666 if (ret) {
1667 fprintf(stderr, "Update of BPF map failed(%d)\n", ret);
1668 exit_with_error(ret);
1669 }
1670 }
1671 }
1672
1673 signal(SIGINT, int_exit);
1674 signal(SIGTERM, int_exit);
1675 signal(SIGABRT, int_exit);
1676
1677 setlocale(LC_ALL, "");
1678
1679 if (!opt_quiet) {
1680 ret = pthread_create(&pt, NULL, poller, NULL);
1681 if (ret)
1682 exit_with_error(ret);
1683 }
1684
1685 prev_time = get_nsecs();
1686 start_time = prev_time;
1687
1688 if (opt_bench == BENCH_RXDROP)
1689 rx_drop_all();
1690 else if (opt_bench == BENCH_TXONLY)
1691 tx_only_all();
1692 else
1693 l2fwd_all();
1694
1695 benchmark_done = true;
1696
1697 if (!opt_quiet)
1698 pthread_join(pt, NULL);
1699
1700 xdpsock_cleanup();
1701
1702 munmap(bufs, NUM_FRAMES * opt_xsk_frame_size);
1703
1704 return 0;
1705 }
1706
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.