1 // SPDX-License-Identifier: GPL-2.0
2 /* Toeplitz test
3  *
4  * 1. Read packets and their rx_hash using PF_PACKET/TPACKET_V3
5  * 2. Compute the rx_hash in software based on the packet contents
6  * 3. Compare the two
7  *
8  * Optionally, either '-C $rx_irq_cpu_list' or '-r $rps_bitmap' may be given.
9  *
10  * If '-C $rx_irq_cpu_list' is given, also
11  *
12  * 4. Identify the cpu on which the packet arrived with PACKET_FANOUT_CPU
13  * 5. Compute the rxqueue that RSS would select based on this rx_hash
14  * 6. Using the $rx_irq_cpu_list map, identify the arriving cpu based on rxq irq
15  * 7. Compare the cpus from 4 and 6
16  *
17  * Else if '-r $rps_bitmap' is given, also
18  *
19  * 4. Identify the cpu on which the packet arrived with PACKET_FANOUT_CPU
20  * 5. Compute the cpu that RPS should select based on rx_hash and $rps_bitmap
21  * 6. Compare the cpus from 4 and 5
22  */
23 
24 #define _GNU_SOURCE
25 
26 #include <arpa/inet.h>
27 #include <errno.h>
28 #include <error.h>
29 #include <fcntl.h>
30 #include <getopt.h>
31 #include <linux/filter.h>
32 #include <linux/if_ether.h>
33 #include <linux/if_packet.h>
34 #include <net/if.h>
35 #include <netdb.h>
36 #include <netinet/ip.h>
37 #include <netinet/ip6.h>
38 #include <netinet/tcp.h>
39 #include <netinet/udp.h>
40 #include <poll.h>
41 #include <stdbool.h>
42 #include <stddef.h>
43 #include <stdint.h>
44 #include <stdio.h>
45 #include <stdlib.h>
46 #include <string.h>
47 #include <sys/mman.h>
48 #include <sys/socket.h>
49 #include <sys/stat.h>
50 #include <sys/sysinfo.h>
51 #include <sys/time.h>
52 #include <sys/types.h>
53 #include <unistd.h>
54 
55 #include "../kselftest.h"
56 
57 #define TOEPLITZ_KEY_MIN_LEN	40
58 #define TOEPLITZ_KEY_MAX_LEN	60
59 
60 #define TOEPLITZ_STR_LEN(K)	(((K) * 3) - 1)	/* hex encoded: AA:BB:CC:...:ZZ */
61 #define TOEPLITZ_STR_MIN_LEN	TOEPLITZ_STR_LEN(TOEPLITZ_KEY_MIN_LEN)
62 #define TOEPLITZ_STR_MAX_LEN	TOEPLITZ_STR_LEN(TOEPLITZ_KEY_MAX_LEN)
63 
64 #define FOUR_TUPLE_MAX_LEN	((sizeof(struct in6_addr) * 2) + (sizeof(uint16_t) * 2))
65 
66 #define RSS_MAX_CPUS (1 << 16)	/* real constraint is PACKET_FANOUT_MAX */
67 
68 #define RPS_MAX_CPUS 16UL	/* must be a power of 2 */
69 
70 /* configuration options (cmdline arguments) */
71 static uint16_t cfg_dport =	8000;
72 static int cfg_family =		AF_INET6;
73 static char *cfg_ifname =	"eth0";
74 static int cfg_num_queues;
75 static int cfg_num_rps_cpus;
76 static bool cfg_sink;
77 static int cfg_type =		SOCK_STREAM;
78 static int cfg_timeout_msec =	1000;
79 static bool cfg_verbose;
80 
81 /* global vars */
82 static int num_cpus;
83 static int ring_block_nr;
84 static int ring_block_sz;
85 
86 /* stats */
87 static int frames_received;
88 static int frames_nohash;
89 static int frames_error;
90 
91 #define log_verbose(args...)	do { if (cfg_verbose) fprintf(stderr, args); } while (0)
92 
93 /* tpacket ring */
94 struct ring_state {
95 	int fd;
96 	char *mmap;
97 	int idx;
98 	int cpu;
99 };
100 
101 static unsigned int rx_irq_cpus[RSS_MAX_CPUS];	/* map from rxq to cpu */
102 static int rps_silo_to_cpu[RPS_MAX_CPUS];
103 static unsigned char toeplitz_key[TOEPLITZ_KEY_MAX_LEN];
104 static struct ring_state rings[RSS_MAX_CPUS];
105 
toeplitz(const unsigned char * four_tuple,const unsigned char * key)106 static inline uint32_t toeplitz(const unsigned char *four_tuple,
107 				const unsigned char *key)
108 {
109 	int i, bit, ret = 0;
110 	uint32_t key32;
111 
112 	key32 = ntohl(*((uint32_t *)key));
113 	key += 4;
114 
115 	for (i = 0; i < FOUR_TUPLE_MAX_LEN; i++) {
116 		for (bit = 7; bit >= 0; bit--) {
117 			if (four_tuple[i] & (1 << bit))
118 				ret ^= key32;
119 
120 			key32 <<= 1;
121 			key32 |= !!(key[0] & (1 << bit));
122 		}
123 		key++;
124 	}
125 
126 	return ret;
127 }
128 
129 /* Compare computed cpu with arrival cpu from packet_fanout_cpu */
verify_rss(uint32_t rx_hash,int cpu)130 static void verify_rss(uint32_t rx_hash, int cpu)
131 {
132 	int queue = rx_hash % cfg_num_queues;
133 
134 	log_verbose(" rxq %d (cpu %d)", queue, rx_irq_cpus[queue]);
135 	if (rx_irq_cpus[queue] != cpu) {
136 		log_verbose(". error: rss cpu mismatch (%d)", cpu);
137 		frames_error++;
138 	}
139 }
140 
verify_rps(uint64_t rx_hash,int cpu)141 static void verify_rps(uint64_t rx_hash, int cpu)
142 {
143 	int silo = (rx_hash * cfg_num_rps_cpus) >> 32;
144 
145 	log_verbose(" silo %d (cpu %d)", silo, rps_silo_to_cpu[silo]);
146 	if (rps_silo_to_cpu[silo] != cpu) {
147 		log_verbose(". error: rps cpu mismatch (%d)", cpu);
148 		frames_error++;
149 	}
150 }
151 
log_rxhash(int cpu,uint32_t rx_hash,const char * addrs,int addr_len)152 static void log_rxhash(int cpu, uint32_t rx_hash,
153 		       const char *addrs, int addr_len)
154 {
155 	char saddr[INET6_ADDRSTRLEN], daddr[INET6_ADDRSTRLEN];
156 	uint16_t *ports;
157 
158 	if (!inet_ntop(cfg_family, addrs, saddr, sizeof(saddr)) ||
159 	    !inet_ntop(cfg_family, addrs + addr_len, daddr, sizeof(daddr)))
160 		error(1, 0, "address parse error");
161 
162 	ports = (void *)addrs + (addr_len * 2);
163 	log_verbose("cpu %d: rx_hash 0x%08x [saddr %s daddr %s sport %02hu dport %02hu]",
164 		    cpu, rx_hash, saddr, daddr,
165 		    ntohs(ports[0]), ntohs(ports[1]));
166 }
167 
168 /* Compare computed rxhash with rxhash received from tpacket_v3 */
verify_rxhash(const char * pkt,uint32_t rx_hash,int cpu)169 static void verify_rxhash(const char *pkt, uint32_t rx_hash, int cpu)
170 {
171 	unsigned char four_tuple[FOUR_TUPLE_MAX_LEN] = {0};
172 	uint32_t rx_hash_sw;
173 	const char *addrs;
174 	int addr_len;
175 
176 	if (cfg_family == AF_INET) {
177 		addr_len = sizeof(struct in_addr);
178 		addrs = pkt + offsetof(struct iphdr, saddr);
179 	} else {
180 		addr_len = sizeof(struct in6_addr);
181 		addrs = pkt + offsetof(struct ip6_hdr, ip6_src);
182 	}
183 
184 	memcpy(four_tuple, addrs, (addr_len * 2) + (sizeof(uint16_t) * 2));
185 	rx_hash_sw = toeplitz(four_tuple, toeplitz_key);
186 
187 	if (cfg_verbose)
188 		log_rxhash(cpu, rx_hash, addrs, addr_len);
189 
190 	if (rx_hash != rx_hash_sw) {
191 		log_verbose(" != expected 0x%x\n", rx_hash_sw);
192 		frames_error++;
193 		return;
194 	}
195 
196 	log_verbose(" OK");
197 	if (cfg_num_queues)
198 		verify_rss(rx_hash, cpu);
199 	else if (cfg_num_rps_cpus)
200 		verify_rps(rx_hash, cpu);
201 	log_verbose("\n");
202 }
203 
recv_frame(const struct ring_state * ring,char * frame)204 static char *recv_frame(const struct ring_state *ring, char *frame)
205 {
206 	struct tpacket3_hdr *hdr = (void *)frame;
207 
208 	if (hdr->hv1.tp_rxhash)
209 		verify_rxhash(frame + hdr->tp_net, hdr->hv1.tp_rxhash,
210 			      ring->cpu);
211 	else
212 		frames_nohash++;
213 
214 	return frame + hdr->tp_next_offset;
215 }
216 
217 /* A single TPACKET_V3 block can hold multiple frames */
recv_block(struct ring_state * ring)218 static bool recv_block(struct ring_state *ring)
219 {
220 	struct tpacket_block_desc *block;
221 	char *frame;
222 	int i;
223 
224 	block = (void *)(ring->mmap + ring->idx * ring_block_sz);
225 	if (!(block->hdr.bh1.block_status & TP_STATUS_USER))
226 		return false;
227 
228 	frame = (char *)block;
229 	frame += block->hdr.bh1.offset_to_first_pkt;
230 
231 	for (i = 0; i < block->hdr.bh1.num_pkts; i++) {
232 		frame = recv_frame(ring, frame);
233 		frames_received++;
234 	}
235 
236 	block->hdr.bh1.block_status = TP_STATUS_KERNEL;
237 	ring->idx = (ring->idx + 1) % ring_block_nr;
238 
239 	return true;
240 }
241 
242 /* simple test: sleep once unconditionally and then process all rings */
process_rings(void)243 static void process_rings(void)
244 {
245 	int i;
246 
247 	usleep(1000 * cfg_timeout_msec);
248 
249 	for (i = 0; i < num_cpus; i++)
250 		do {} while (recv_block(&rings[i]));
251 
252 	fprintf(stderr, "count: pass=%u nohash=%u fail=%u\n",
253 		frames_received - frames_nohash - frames_error,
254 		frames_nohash, frames_error);
255 }
256 
setup_ring(int fd)257 static char *setup_ring(int fd)
258 {
259 	struct tpacket_req3 req3 = {0};
260 	void *ring;
261 
262 	req3.tp_retire_blk_tov = cfg_timeout_msec / 8;
263 	req3.tp_feature_req_word = TP_FT_REQ_FILL_RXHASH;
264 
265 	req3.tp_frame_size = 2048;
266 	req3.tp_frame_nr = 1 << 10;
267 	req3.tp_block_nr = 16;
268 
269 	req3.tp_block_size = req3.tp_frame_size * req3.tp_frame_nr;
270 	req3.tp_block_size /= req3.tp_block_nr;
271 
272 	if (setsockopt(fd, SOL_PACKET, PACKET_RX_RING, &req3, sizeof(req3)))
273 		error(1, errno, "setsockopt PACKET_RX_RING");
274 
275 	ring_block_sz = req3.tp_block_size;
276 	ring_block_nr = req3.tp_block_nr;
277 
278 	ring = mmap(0, req3.tp_block_size * req3.tp_block_nr,
279 		    PROT_READ | PROT_WRITE,
280 		    MAP_SHARED | MAP_LOCKED | MAP_POPULATE, fd, 0);
281 	if (ring == MAP_FAILED)
282 		error(1, 0, "mmap failed");
283 
284 	return ring;
285 }
286 
__set_filter(int fd,int off_proto,uint8_t proto,int off_dport)287 static void __set_filter(int fd, int off_proto, uint8_t proto, int off_dport)
288 {
289 	struct sock_filter filter[] = {
290 		BPF_STMT(BPF_LD  + BPF_B   + BPF_ABS, SKF_AD_OFF + SKF_AD_PKTTYPE),
291 		BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, PACKET_HOST, 0, 4),
292 		BPF_STMT(BPF_LD  + BPF_B   + BPF_ABS, off_proto),
293 		BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, proto, 0, 2),
294 		BPF_STMT(BPF_LD  + BPF_H   + BPF_ABS, off_dport),
295 		BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, cfg_dport, 1, 0),
296 		BPF_STMT(BPF_RET + BPF_K, 0),
297 		BPF_STMT(BPF_RET + BPF_K, 0xFFFF),
298 	};
299 	struct sock_fprog prog = {};
300 
301 	prog.filter = filter;
302 	prog.len = ARRAY_SIZE(filter);
303 	if (setsockopt(fd, SOL_SOCKET, SO_ATTACH_FILTER, &prog, sizeof(prog)))
304 		error(1, errno, "setsockopt filter");
305 }
306 
307 /* filter on transport protocol and destination port */
set_filter(int fd)308 static void set_filter(int fd)
309 {
310 	const int off_dport = offsetof(struct tcphdr, dest);	/* same for udp */
311 	uint8_t proto;
312 
313 	proto = cfg_type == SOCK_STREAM ? IPPROTO_TCP : IPPROTO_UDP;
314 	if (cfg_family == AF_INET)
315 		__set_filter(fd, offsetof(struct iphdr, protocol), proto,
316 			     sizeof(struct iphdr) + off_dport);
317 	else
318 		__set_filter(fd, offsetof(struct ip6_hdr, ip6_nxt), proto,
319 			     sizeof(struct ip6_hdr) + off_dport);
320 }
321 
322 /* drop everything: used temporarily during setup */
set_filter_null(int fd)323 static void set_filter_null(int fd)
324 {
325 	struct sock_filter filter[] = {
326 		BPF_STMT(BPF_RET + BPF_K, 0),
327 	};
328 	struct sock_fprog prog = {};
329 
330 	prog.filter = filter;
331 	prog.len = ARRAY_SIZE(filter);
332 	if (setsockopt(fd, SOL_SOCKET, SO_ATTACH_FILTER, &prog, sizeof(prog)))
333 		error(1, errno, "setsockopt filter");
334 }
335 
create_ring(char ** ring)336 static int create_ring(char **ring)
337 {
338 	struct fanout_args args = {
339 		.id = 1,
340 		.type_flags = PACKET_FANOUT_CPU,
341 		.max_num_members = RSS_MAX_CPUS
342 	};
343 	struct sockaddr_ll ll = { 0 };
344 	int fd, val;
345 
346 	fd = socket(PF_PACKET, SOCK_DGRAM, 0);
347 	if (fd == -1)
348 		error(1, errno, "socket creation failed");
349 
350 	val = TPACKET_V3;
351 	if (setsockopt(fd, SOL_PACKET, PACKET_VERSION, &val, sizeof(val)))
352 		error(1, errno, "setsockopt PACKET_VERSION");
353 	*ring = setup_ring(fd);
354 
355 	/* block packets until all rings are added to the fanout group:
356 	 * else packets can arrive during setup and get misclassified
357 	 */
358 	set_filter_null(fd);
359 
360 	ll.sll_family = AF_PACKET;
361 	ll.sll_ifindex = if_nametoindex(cfg_ifname);
362 	ll.sll_protocol = cfg_family == AF_INET ? htons(ETH_P_IP) :
363 						  htons(ETH_P_IPV6);
364 	if (bind(fd, (void *)&ll, sizeof(ll)))
365 		error(1, errno, "bind");
366 
367 	/* must come after bind: verifies all programs in group match */
368 	if (setsockopt(fd, SOL_PACKET, PACKET_FANOUT, &args, sizeof(args))) {
369 		/* on failure, retry using old API if that is sufficient:
370 		 * it has a hard limit of 256 sockets, so only try if
371 		 * (a) only testing rxhash, not RSS or (b) <= 256 cpus.
372 		 * in this API, the third argument is left implicit.
373 		 */
374 		if (cfg_num_queues || num_cpus > 256 ||
375 		    setsockopt(fd, SOL_PACKET, PACKET_FANOUT,
376 			       &args, sizeof(uint32_t)))
377 			error(1, errno, "setsockopt PACKET_FANOUT cpu");
378 	}
379 
380 	return fd;
381 }
382 
383 /* setup inet(6) socket to blackhole the test traffic, if arg '-s' */
setup_sink(void)384 static int setup_sink(void)
385 {
386 	int fd, val;
387 
388 	fd = socket(cfg_family, cfg_type, 0);
389 	if (fd == -1)
390 		error(1, errno, "socket %d.%d", cfg_family, cfg_type);
391 
392 	val = 1 << 20;
393 	if (setsockopt(fd, SOL_SOCKET, SO_RCVBUFFORCE, &val, sizeof(val)))
394 		error(1, errno, "setsockopt rcvbuf");
395 
396 	return fd;
397 }
398 
setup_rings(void)399 static void setup_rings(void)
400 {
401 	int i;
402 
403 	for (i = 0; i < num_cpus; i++) {
404 		rings[i].cpu = i;
405 		rings[i].fd = create_ring(&rings[i].mmap);
406 	}
407 
408 	/* accept packets once all rings in the fanout group are up */
409 	for (i = 0; i < num_cpus; i++)
410 		set_filter(rings[i].fd);
411 }
412 
cleanup_rings(void)413 static void cleanup_rings(void)
414 {
415 	int i;
416 
417 	for (i = 0; i < num_cpus; i++) {
418 		if (munmap(rings[i].mmap, ring_block_nr * ring_block_sz))
419 			error(1, errno, "munmap");
420 		if (close(rings[i].fd))
421 			error(1, errno, "close");
422 	}
423 }
424 
parse_cpulist(const char * arg)425 static void parse_cpulist(const char *arg)
426 {
427 	do {
428 		rx_irq_cpus[cfg_num_queues++] = strtol(arg, NULL, 10);
429 
430 		arg = strchr(arg, ',');
431 		if (!arg)
432 			break;
433 		arg++;			// skip ','
434 	} while (1);
435 }
436 
show_cpulist(void)437 static void show_cpulist(void)
438 {
439 	int i;
440 
441 	for (i = 0; i < cfg_num_queues; i++)
442 		fprintf(stderr, "rxq %d: cpu %d\n", i, rx_irq_cpus[i]);
443 }
444 
show_silos(void)445 static void show_silos(void)
446 {
447 	int i;
448 
449 	for (i = 0; i < cfg_num_rps_cpus; i++)
450 		fprintf(stderr, "silo %d: cpu %d\n", i, rps_silo_to_cpu[i]);
451 }
452 
parse_toeplitz_key(const char * str,int slen,unsigned char * key)453 static void parse_toeplitz_key(const char *str, int slen, unsigned char *key)
454 {
455 	int i, ret, off;
456 
457 	if (slen < TOEPLITZ_STR_MIN_LEN ||
458 	    slen > TOEPLITZ_STR_MAX_LEN + 1)
459 		error(1, 0, "invalid toeplitz key");
460 
461 	for (i = 0, off = 0; off < slen; i++, off += 3) {
462 		ret = sscanf(str + off, "%hhx", &key[i]);
463 		if (ret != 1)
464 			error(1, 0, "key parse error at %d off %d len %d",
465 			      i, off, slen);
466 	}
467 }
468 
parse_rps_bitmap(const char * arg)469 static void parse_rps_bitmap(const char *arg)
470 {
471 	unsigned long bitmap;
472 	int i;
473 
474 	bitmap = strtoul(arg, NULL, 0);
475 
476 	if (bitmap & ~(RPS_MAX_CPUS - 1))
477 		error(1, 0, "rps bitmap 0x%lx out of bounds 0..%lu",
478 		      bitmap, RPS_MAX_CPUS - 1);
479 
480 	for (i = 0; i < RPS_MAX_CPUS; i++)
481 		if (bitmap & 1UL << i)
482 			rps_silo_to_cpu[cfg_num_rps_cpus++] = i;
483 }
484 
parse_opts(int argc,char ** argv)485 static void parse_opts(int argc, char **argv)
486 {
487 	static struct option long_options[] = {
488 	    {"dport",	required_argument, 0, 'd'},
489 	    {"cpus",	required_argument, 0, 'C'},
490 	    {"key",	required_argument, 0, 'k'},
491 	    {"iface",	required_argument, 0, 'i'},
492 	    {"ipv4",	no_argument, 0, '4'},
493 	    {"ipv6",	no_argument, 0, '6'},
494 	    {"sink",	no_argument, 0, 's'},
495 	    {"tcp",	no_argument, 0, 't'},
496 	    {"timeout",	required_argument, 0, 'T'},
497 	    {"udp",	no_argument, 0, 'u'},
498 	    {"verbose",	no_argument, 0, 'v'},
499 	    {"rps",	required_argument, 0, 'r'},
500 	    {0, 0, 0, 0}
501 	};
502 	bool have_toeplitz = false;
503 	int index, c;
504 
505 	while ((c = getopt_long(argc, argv, "46C:d:i:k:r:stT:uv", long_options, &index)) != -1) {
506 		switch (c) {
507 		case '4':
508 			cfg_family = AF_INET;
509 			break;
510 		case '6':
511 			cfg_family = AF_INET6;
512 			break;
513 		case 'C':
514 			parse_cpulist(optarg);
515 			break;
516 		case 'd':
517 			cfg_dport = strtol(optarg, NULL, 0);
518 			break;
519 		case 'i':
520 			cfg_ifname = optarg;
521 			break;
522 		case 'k':
523 			parse_toeplitz_key(optarg, strlen(optarg),
524 					   toeplitz_key);
525 			have_toeplitz = true;
526 			break;
527 		case 'r':
528 			parse_rps_bitmap(optarg);
529 			break;
530 		case 's':
531 			cfg_sink = true;
532 			break;
533 		case 't':
534 			cfg_type = SOCK_STREAM;
535 			break;
536 		case 'T':
537 			cfg_timeout_msec = strtol(optarg, NULL, 0);
538 			break;
539 		case 'u':
540 			cfg_type = SOCK_DGRAM;
541 			break;
542 		case 'v':
543 			cfg_verbose = true;
544 			break;
545 
546 		default:
547 			error(1, 0, "unknown option %c", optopt);
548 			break;
549 		}
550 	}
551 
552 	if (!have_toeplitz)
553 		error(1, 0, "Must supply rss key ('-k')");
554 
555 	num_cpus = get_nprocs();
556 	if (num_cpus > RSS_MAX_CPUS)
557 		error(1, 0, "increase RSS_MAX_CPUS");
558 
559 	if (cfg_num_queues && cfg_num_rps_cpus)
560 		error(1, 0,
561 		      "Can't supply both RSS cpus ('-C') and RPS map ('-r')");
562 	if (cfg_verbose) {
563 		show_cpulist();
564 		show_silos();
565 	}
566 }
567 
main(int argc,char ** argv)568 int main(int argc, char **argv)
569 {
570 	const int min_tests = 10;
571 	int fd_sink = -1;
572 
573 	parse_opts(argc, argv);
574 
575 	if (cfg_sink)
576 		fd_sink = setup_sink();
577 
578 	setup_rings();
579 	process_rings();
580 	cleanup_rings();
581 
582 	if (cfg_sink && close(fd_sink))
583 		error(1, errno, "close sink");
584 
585 	if (frames_received - frames_nohash < min_tests)
586 		error(1, 0, "too few frames for verification");
587 
588 	return frames_error;
589 }
590