xref: /linux/tools/testing/selftests/bpf/progs/xdp_lb_bench.c (revision b2128290c29902315e632ea59e0504d6bc9e9b42)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2026 Meta Platforms, Inc. and affiliates. */
3 
4 #include <stddef.h>
5 #include <stdbool.h>
6 #include <linux/bpf.h>
7 #include <linux/if_ether.h>
8 #include <linux/ip.h>
9 #include <linux/ipv6.h>
10 #include <linux/in.h>
11 #include <linux/tcp.h>
12 #include <linux/udp.h>
13 #include <bpf/bpf_helpers.h>
14 #include <bpf/bpf_endian.h>
15 #include "bpf_compiler.h"
16 #include "xdp_lb_bench_common.h"
17 #include "bench_bpf_timing.bpf.h"
18 
19 #ifndef IPPROTO_FRAGMENT
20 #define IPPROTO_FRAGMENT 44
21 #endif
22 
23 /* jhash helpers */
24 
25 static inline __u32 rol32(__u32 word, unsigned int shift)
26 {
27 	return (word << shift) | (word >> ((-shift) & 31));
28 }
29 
30 #define __jhash_mix(a, b, c)			\
31 {						\
32 	a -= c;  a ^= rol32(c, 4);  c += b;	\
33 	b -= a;  b ^= rol32(a, 6);  a += c;	\
34 	c -= b;  c ^= rol32(b, 8);  b += a;	\
35 	a -= c;  a ^= rol32(c, 16); c += b;	\
36 	b -= a;  b ^= rol32(a, 19); a += c;	\
37 	c -= b;  c ^= rol32(b, 4);  b += a;	\
38 }
39 
40 #define __jhash_final(a, b, c)			\
41 {						\
42 	c ^= b; c -= rol32(b, 14);		\
43 	a ^= c; a -= rol32(c, 11);		\
44 	b ^= a; b -= rol32(a, 25);		\
45 	c ^= b; c -= rol32(b, 16);		\
46 	a ^= c; a -= rol32(c, 4);		\
47 	b ^= a; b -= rol32(a, 14);		\
48 	c ^= b; c -= rol32(b, 24);		\
49 }
50 
51 #define JHASH_INITVAL 0xdeadbeef
52 
53 static inline __u32 __jhash_nwords(__u32 a, __u32 b, __u32 c, __u32 initval)
54 {
55 	a += initval;
56 	b += initval;
57 	c += initval;
58 	__jhash_final(a, b, c);
59 	return c;
60 }
61 
62 static inline __u32 jhash_2words(__u32 a, __u32 b, __u32 initval)
63 {
64 	return __jhash_nwords(a, b, 0, initval + JHASH_INITVAL + (2 << 2));
65 }
66 
67 static inline __u32 jhash2_4words(const __u32 *k, __u32 initval)
68 {
69 	__u32 a, b, c;
70 
71 	a = b = c = JHASH_INITVAL + (4 << 2) + initval;
72 
73 	a += k[0]; b += k[1]; c += k[2];
74 	__jhash_mix(a, b, c);
75 
76 	a += k[3];
77 	__jhash_final(a, b, c);
78 
79 	return c;
80 }
81 
82 static __always_inline void ipv4_csum(struct iphdr *iph)
83 {
84 	__u16 *next_iph = (__u16 *)iph;
85 	__u32 csum = 0;
86 	int i;
87 
88 	__pragma_loop_unroll_full
89 	for (i = 0; i < (int)(sizeof(*iph) >> 1); i++)
90 		csum += *next_iph++;
91 
92 	csum = (csum & 0xffff) + (csum >> 16);
93 	csum = (csum & 0xffff) + (csum >> 16);
94 	iph->check = ~csum;
95 }
96 
97 struct {
98 	__uint(type, BPF_MAP_TYPE_HASH);
99 	__uint(max_entries, 64);
100 	__type(key, struct vip_definition);
101 	__type(value, struct vip_meta);
102 } vip_map SEC(".maps");
103 
104 struct lru_inner_map {
105 	__uint(type, BPF_MAP_TYPE_LRU_HASH);
106 	__type(key, struct flow_key);
107 	__type(value, struct real_pos_lru);
108 	__uint(max_entries, DEFAULT_LRU_SIZE);
109 } lru_inner SEC(".maps");
110 
111 struct {
112 	__uint(type, BPF_MAP_TYPE_ARRAY_OF_MAPS);
113 	__type(key, __u32);
114 	__type(value, __u32);
115 	__uint(max_entries, BENCH_NR_CPUS);
116 	__array(values, struct lru_inner_map);
117 } lru_mapping SEC(".maps");
118 
119 struct {
120 	__uint(type, BPF_MAP_TYPE_ARRAY);
121 	__uint(max_entries, CH_RINGS_SIZE);
122 	__type(key, __u32);
123 	__type(value, __u32);
124 } ch_rings SEC(".maps");
125 
126 struct {
127 	__uint(type, BPF_MAP_TYPE_ARRAY);
128 	__uint(max_entries, MAX_REALS);
129 	__type(key, __u32);
130 	__type(value, struct real_definition);
131 } reals SEC(".maps");
132 
133 struct {
134 	__uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
135 	__uint(max_entries, STATS_SIZE);
136 	__type(key, __u32);
137 	__type(value, struct lb_stats);
138 } stats SEC(".maps");
139 
140 struct {
141 	__uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
142 	__uint(max_entries, MAX_REALS);
143 	__type(key, __u32);
144 	__type(value, struct lb_stats);
145 } reals_stats SEC(".maps");
146 
147 struct {
148 	__uint(type, BPF_MAP_TYPE_ARRAY);
149 	__uint(max_entries, 1);
150 	__type(key, __u32);
151 	__type(value, struct ctl_value);
152 } ctl_array SEC(".maps");
153 
154 struct {
155 	__uint(type, BPF_MAP_TYPE_ARRAY);
156 	__uint(max_entries, 1);
157 	__type(key, __u32);
158 	__type(value, struct vip_definition);
159 } vip_miss_stats SEC(".maps");
160 
161 struct {
162 	__uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
163 	__uint(max_entries, MAX_REALS);
164 	__type(key, __u32);
165 	__type(value, __u32);
166 } lru_miss_stats SEC(".maps");
167 
168 volatile __u32 flow_mask;
169 volatile __u32 cold_lru;
170 __u32 batch_gen;
171 
172 /*
173  * old_eth MUST be read BEFORE writing the outer header because
174  * bpf_xdp_adjust_head makes them overlap.
175  */
176 static __always_inline int encap_v4(struct xdp_md *xdp, __be32 saddr, __be32 daddr,
177 				    __u16 payload_len, const __u8 *dst_mac)
178 {
179 	struct ethhdr *new_eth, *old_eth;
180 	void *data, *data_end;
181 	struct iphdr *iph;
182 
183 	if (bpf_xdp_adjust_head(xdp, -(int)sizeof(struct iphdr)))
184 		return -1;
185 
186 	data     = (void *)(long)xdp->data;
187 	data_end = (void *)(long)xdp->data_end;
188 
189 	new_eth = data;
190 	iph     = data + sizeof(struct ethhdr);
191 	old_eth = data + sizeof(struct iphdr);
192 
193 	if (new_eth + 1 > data_end || old_eth + 1 > data_end || iph + 1 > data_end)
194 		return -1;
195 
196 	__builtin_memcpy(new_eth->h_source, old_eth->h_dest, sizeof(new_eth->h_source));
197 	__builtin_memcpy(new_eth->h_dest, dst_mac, sizeof(new_eth->h_dest));
198 	new_eth->h_proto = bpf_htons(ETH_P_IP);
199 
200 	__builtin_memset(iph, 0, sizeof(*iph));
201 	iph->version  = 4;
202 	iph->ihl      = sizeof(*iph) >> 2;
203 	iph->protocol = IPPROTO_IPIP;
204 	iph->tot_len  = bpf_htons(payload_len + sizeof(*iph));
205 	iph->ttl      = 64;
206 	iph->saddr    = saddr;
207 	iph->daddr    = daddr;
208 	ipv4_csum(iph);
209 
210 	return 0;
211 }
212 
213 static __always_inline int encap_v6(struct xdp_md *xdp, const __be32 saddr[4],
214 				    const __be32 daddr[4], __u8 nexthdr, __u16 payload_len,
215 				    const __u8 *dst_mac)
216 {
217 	struct ethhdr *new_eth, *old_eth;
218 	void *data, *data_end;
219 	struct ipv6hdr *ip6h;
220 
221 	if (bpf_xdp_adjust_head(xdp, -(int)sizeof(struct ipv6hdr)))
222 		return -1;
223 
224 	data     = (void *)(long)xdp->data;
225 	data_end = (void *)(long)xdp->data_end;
226 
227 	new_eth = data;
228 	ip6h    = data + sizeof(struct ethhdr);
229 	old_eth = data + sizeof(struct ipv6hdr);
230 
231 	if (new_eth + 1 > data_end || old_eth + 1 > data_end || ip6h + 1 > data_end)
232 		return -1;
233 
234 	__builtin_memcpy(new_eth->h_source, old_eth->h_dest, sizeof(new_eth->h_source));
235 	__builtin_memcpy(new_eth->h_dest, dst_mac, sizeof(new_eth->h_dest));
236 	new_eth->h_proto = bpf_htons(ETH_P_IPV6);
237 
238 	__builtin_memset(ip6h, 0, sizeof(*ip6h));
239 	ip6h->version     = 6;
240 	ip6h->nexthdr     = nexthdr;
241 	ip6h->payload_len = bpf_htons(payload_len);
242 	ip6h->hop_limit   = 64;
243 	__builtin_memcpy(&ip6h->saddr, saddr, sizeof(ip6h->saddr));
244 	__builtin_memcpy(&ip6h->daddr, daddr, sizeof(ip6h->daddr));
245 
246 	return 0;
247 }
248 
249 static __always_inline void update_stats(void *map, __u32 key, __u16 bytes)
250 {
251 	struct lb_stats *st = bpf_map_lookup_elem(map, &key);
252 
253 	if (st) {
254 		st->v1 += 1;
255 		st->v2 += bytes;
256 	}
257 }
258 
259 static __always_inline void count_action(int action)
260 {
261 	struct lb_stats *st;
262 	__u32 key;
263 
264 	if (action == XDP_TX)
265 		key = STATS_XDP_TX;
266 	else if (action == XDP_PASS)
267 		key = STATS_XDP_PASS;
268 	else
269 		key = STATS_XDP_DROP;
270 
271 	st = bpf_map_lookup_elem(&stats, &key);
272 	if (st)
273 		st->v1 += 1;
274 }
275 
276 static __always_inline bool is_under_flood(void)
277 {
278 	__u32 key = STATS_NEW_CONN;
279 	struct lb_stats *conn_st = bpf_map_lookup_elem(&stats, &key);
280 	__u64 cur_time;
281 
282 	if (!conn_st)
283 		return true;
284 
285 	cur_time = bpf_ktime_get_ns();
286 	if ((cur_time - conn_st->v2) > ONE_SEC) {
287 		conn_st->v1 = 1;
288 		conn_st->v2 = cur_time;
289 	} else {
290 		conn_st->v1 += 1;
291 		if (conn_st->v1 > MAX_CONN_RATE)
292 			return true;
293 	}
294 	return false;
295 }
296 
297 static __always_inline struct real_definition *connection_table_lookup(void *lru_map,
298 								       struct flow_key *flow,
299 								       __u32 *out_pos)
300 {
301 	struct real_pos_lru *dst_lru;
302 	struct real_definition *real;
303 	__u32 key;
304 
305 	dst_lru = bpf_map_lookup_elem(lru_map, flow);
306 	if (!dst_lru)
307 		return NULL;
308 
309 	/* UDP connections use atime-based timeout instead of FIN/RST */
310 	if (flow->proto == IPPROTO_UDP) {
311 		__u64 cur_time = bpf_ktime_get_ns();
312 
313 		if (cur_time - dst_lru->atime > LRU_UDP_TIMEOUT)
314 			return NULL;
315 		dst_lru->atime = cur_time;
316 	}
317 
318 	key = dst_lru->pos;
319 	*out_pos = key;
320 	real = bpf_map_lookup_elem(&reals, &key);
321 	return real;
322 }
323 
324 static __always_inline bool get_packet_dst(struct real_definition **real, struct flow_key *flow,
325 					   struct vip_meta *vip_info, bool is_v6, void *lru_map,
326 					   bool is_rst, __u32 *out_pos)
327 {
328 	bool under_flood;
329 	__u32 hash, ch_key;
330 	__u32 *ch_val;
331 	__u32 real_pos;
332 
333 	under_flood = is_under_flood();
334 
335 	if (is_v6) {
336 		__u32 src_hash = jhash2_4words((__u32 *)flow->srcv6, MAX_VIPS);
337 
338 		hash = jhash_2words(src_hash, flow->ports, CH_RING_SIZE);
339 	} else {
340 		hash = jhash_2words(flow->src, flow->ports, CH_RING_SIZE);
341 	}
342 
343 	ch_key = CH_RING_SIZE * vip_info->vip_num + hash % CH_RING_SIZE;
344 	ch_val = bpf_map_lookup_elem(&ch_rings, &ch_key);
345 	if (!ch_val)
346 		return false;
347 	real_pos = *ch_val;
348 
349 	*real = bpf_map_lookup_elem(&reals, &real_pos);
350 	if (!(*real))
351 		return false;
352 
353 	if (!(vip_info->flags & F_LRU_BYPASS) && !under_flood && !is_rst) {
354 		struct real_pos_lru new_lru = { .pos = real_pos };
355 
356 		if (flow->proto == IPPROTO_UDP)
357 			new_lru.atime = bpf_ktime_get_ns();
358 		bpf_map_update_elem(lru_map, flow, &new_lru, BPF_ANY);
359 	}
360 
361 	*out_pos = real_pos;
362 	return true;
363 }
364 
365 static __always_inline void update_vip_lru_miss_stats(struct vip_definition *vip, bool is_v6,
366 						      __u32 real_idx)
367 {
368 	struct vip_definition *miss_vip;
369 	__u32 key = 0;
370 	__u32 *cnt;
371 
372 	miss_vip = bpf_map_lookup_elem(&vip_miss_stats, &key);
373 	if (!miss_vip)
374 		return;
375 
376 	if (is_v6) {
377 		if (miss_vip->vipv6[0] != vip->vipv6[0] || miss_vip->vipv6[1] != vip->vipv6[1] ||
378 		    miss_vip->vipv6[2] != vip->vipv6[2] || miss_vip->vipv6[3] != vip->vipv6[3])
379 			return;
380 	} else {
381 		if (miss_vip->vip != vip->vip)
382 			return;
383 	}
384 
385 	if (miss_vip->port != vip->port || miss_vip->proto != vip->proto)
386 		return;
387 
388 	cnt = bpf_map_lookup_elem(&lru_miss_stats, &real_idx);
389 	if (cnt)
390 		*cnt += 1;
391 }
392 
393 static __noinline int process_packet(struct xdp_md *xdp)
394 {
395 	void *data     = (void *)(long)xdp->data;
396 	void *data_end = (void *)(long)xdp->data_end;
397 	struct ethhdr *eth = data;
398 	struct real_definition *dst = NULL;
399 	struct vip_definition vip_def = {};
400 	struct ctl_value *cval;
401 	struct flow_key flow = {};
402 	struct vip_meta *vip_info;
403 	struct lb_stats *data_stats;
404 	struct udphdr *uh;
405 	__be32 tnl_src[4];
406 	void *lru_map;
407 	void *l4;
408 	__u16 payload_len;
409 	__u32 real_pos = 0, cpu_num, key;
410 	__u8 proto;
411 	int action = XDP_DROP;
412 	bool is_v6, is_syn = false, is_rst = false;
413 
414 	if (eth + 1 > data_end)
415 		goto out;
416 
417 	if (eth->h_proto == bpf_htons(ETH_P_IPV6)) {
418 		is_v6 = true;
419 	} else if (eth->h_proto == bpf_htons(ETH_P_IP)) {
420 		is_v6 = false;
421 	} else {
422 		action = XDP_PASS;
423 		goto out;
424 	}
425 
426 	if (is_v6) {
427 		struct ipv6hdr *ip6h = (void *)(eth + 1);
428 
429 		if (ip6h + 1 > data_end)
430 			goto out;
431 		if (ip6h->nexthdr == IPPROTO_FRAGMENT)
432 			goto out;
433 
434 		payload_len = sizeof(struct ipv6hdr) + bpf_ntohs(ip6h->payload_len);
435 		proto = ip6h->nexthdr;
436 
437 		__builtin_memcpy(flow.srcv6, &ip6h->saddr, sizeof(flow.srcv6));
438 		__builtin_memcpy(flow.dstv6, &ip6h->daddr, sizeof(flow.dstv6));
439 		__builtin_memcpy(vip_def.vipv6, &ip6h->daddr, sizeof(vip_def.vipv6));
440 		l4 = (void *)(ip6h + 1);
441 	} else {
442 		struct iphdr *iph = (void *)(eth + 1);
443 
444 		if (iph + 1 > data_end)
445 			goto out;
446 		if (iph->ihl != 5)
447 			goto out;
448 		if (iph->frag_off & bpf_htons(PCKT_FRAGMENTED))
449 			goto out;
450 
451 		payload_len = bpf_ntohs(iph->tot_len);
452 		proto = iph->protocol;
453 
454 		flow.src    = iph->saddr;
455 		flow.dst    = iph->daddr;
456 		vip_def.vip = iph->daddr;
457 		l4 = (void *)(iph + 1);
458 	}
459 
460 	/* TCP and UDP share the same port layout at offset 0 */
461 	if (proto != IPPROTO_TCP && proto != IPPROTO_UDP) {
462 		action = XDP_PASS;
463 		goto out;
464 	}
465 
466 	uh = l4;
467 	if ((void *)(uh + 1) > data_end)
468 		goto out;
469 	flow.port16[0] = uh->source;
470 	flow.port16[1] = uh->dest;
471 
472 	if (proto == IPPROTO_TCP) {
473 		struct tcphdr *th = l4;
474 
475 		if ((void *)(th + 1) > data_end)
476 			goto out;
477 		is_syn = th->syn;
478 		is_rst = th->rst;
479 	}
480 
481 	flow.proto    = proto;
482 	vip_def.port  = flow.port16[1];
483 	vip_def.proto = proto;
484 
485 	vip_info = bpf_map_lookup_elem(&vip_map, &vip_def);
486 	if (!vip_info) {
487 		action = XDP_PASS;
488 		goto out;
489 	}
490 
491 	key = STATS_LRU;
492 	data_stats = bpf_map_lookup_elem(&stats, &key);
493 	if (!data_stats)
494 		goto out;
495 	data_stats->v1 += 1;
496 
497 	cpu_num = bpf_get_smp_processor_id();
498 	lru_map = bpf_map_lookup_elem(&lru_mapping, &cpu_num);
499 	if (!lru_map)
500 		goto out;
501 
502 	if (!(vip_info->flags & F_LRU_BYPASS) && !is_syn)
503 		dst = connection_table_lookup(lru_map, &flow, &real_pos);
504 
505 	if (!dst) {
506 		if (flow.proto == IPPROTO_TCP) {
507 			struct lb_stats *miss_st;
508 
509 			key = STATS_LRU_MISS;
510 			miss_st = bpf_map_lookup_elem(&stats, &key);
511 			if (miss_st)
512 				miss_st->v1 += 1;
513 		}
514 
515 		if (!get_packet_dst(&dst, &flow, vip_info, is_v6, lru_map, is_rst, &real_pos))
516 			goto out;
517 
518 		update_vip_lru_miss_stats(&vip_def, is_v6, real_pos);
519 		data_stats->v2 += 1;
520 	}
521 
522 	key = 0;
523 	cval = bpf_map_lookup_elem(&ctl_array, &key);
524 	if (!cval)
525 		goto out;
526 
527 	update_stats(&stats, vip_info->vip_num, payload_len);
528 	update_stats(&reals_stats, real_pos, payload_len);
529 
530 	if (is_v6) {
531 		create_encap_ipv6_src(flow.port16[0], flow.srcv6[0], tnl_src);
532 		if (encap_v6(xdp, tnl_src, dst->dstv6, IPPROTO_IPV6, payload_len, cval->mac))
533 			goto out;
534 	} else if (dst->flags & F_IPV6) {
535 		create_encap_ipv6_src(flow.port16[0], flow.src, tnl_src);
536 		if (encap_v6(xdp, tnl_src, dst->dstv6, IPPROTO_IPIP, payload_len, cval->mac))
537 			goto out;
538 	} else {
539 		if (encap_v4(xdp, create_encap_ipv4_src(flow.port16[0], flow.src), dst->dst,
540 			     payload_len, cval->mac))
541 			goto out;
542 	}
543 
544 	action = XDP_TX;
545 
546 out:
547 	count_action(action);
548 	return action;
549 }
550 
551 static __always_inline int strip_encap(struct xdp_md *xdp, const struct ethhdr *saved_eth)
552 {
553 	void *data = (void *)(long)xdp->data;
554 	void *data_end = (void *)(long)xdp->data_end;
555 	struct ethhdr *eth = data;
556 	int hdr_sz;
557 
558 	if (eth + 1 > data_end)
559 		return -1;
560 
561 	hdr_sz = (eth->h_proto == bpf_htons(ETH_P_IPV6)) ? (int)sizeof(struct ipv6hdr)
562 							 : (int)sizeof(struct iphdr);
563 
564 	if (bpf_xdp_adjust_head(xdp, hdr_sz))
565 		return -1;
566 
567 	data     = (void *)(long)xdp->data;
568 	data_end = (void *)(long)xdp->data_end;
569 	eth      = data;
570 
571 	if (eth + 1 > data_end)
572 		return -1;
573 
574 	__builtin_memcpy(eth, saved_eth, sizeof(*saved_eth));
575 	return 0;
576 }
577 
578 static __always_inline void randomize_src(struct xdp_md *xdp, int saddr_off, __u32 *rand_state)
579 {
580 	void *data     = (void *)(long)xdp->data;
581 	void *data_end = (void *)(long)xdp->data_end;
582 	__u32 *saddr   = data + saddr_off;
583 
584 	*rand_state ^= *rand_state << 13;
585 	*rand_state ^= *rand_state >> 17;
586 	*rand_state ^= *rand_state << 5;
587 
588 	if ((void *)(saddr + 1) <= data_end)
589 		*saddr = *rand_state & flow_mask;
590 }
591 
592 SEC("xdp")
593 int xdp_lb_bench(struct xdp_md *xdp)
594 {
595 	void *data     = (void *)(long)xdp->data;
596 	void *data_end = (void *)(long)xdp->data_end;
597 	struct ethhdr *eth = data;
598 	struct ethhdr saved_eth;
599 	__u32 rand_state = 0;
600 	__u32 batch_hash = 0;
601 	int saddr_off = 0;
602 	bool is_v6;
603 
604 	if (eth + 1 > data_end)
605 		return XDP_DROP;
606 
607 	__builtin_memcpy(&saved_eth, eth, sizeof(saved_eth));
608 
609 	is_v6 = (saved_eth.h_proto == bpf_htons(ETH_P_IPV6));
610 
611 	saddr_off = sizeof(struct ethhdr) + (is_v6 ? offsetof(struct ipv6hdr, saddr) :
612 					     offsetof(struct iphdr, saddr));
613 
614 	if (flow_mask)
615 		rand_state = bpf_get_prandom_u32() | 1;
616 
617 	if (cold_lru) {
618 		__u32 *saddr = data + saddr_off;
619 
620 		batch_gen++;
621 		batch_hash = (batch_gen + bpf_get_smp_processor_id()) * KNUTH_HASH_MULT;
622 		if ((void *)(saddr + 1) <= data_end)
623 			*saddr ^= batch_hash;
624 	}
625 
626 	return BENCH_BPF_LOOP(
627 		process_packet(xdp),
628 		({
629 			if (__bench_result == XDP_TX) {
630 				if (strip_encap(xdp, &saved_eth))
631 					return XDP_DROP;
632 				if (rand_state)
633 					randomize_src(xdp, saddr_off, &rand_state);
634 			}
635 			if (cold_lru) {
636 				void *d = (void *)(long)xdp->data;
637 				void *de = (void *)(long)xdp->data_end;
638 				__u32 *__sa = d + saddr_off;
639 
640 				if ((void *)(__sa + 1) <= de)
641 					*__sa ^= batch_hash;
642 			}
643 		})
644 	);
645 }
646 
647 char _license[] SEC("license") = "GPL";
648