xref: /linux/net/ipv4/udp_offload.c (revision 87579b8cda9ec6ecba8327c59d8505d3b83fda98)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  *	IPV4 GSO/GRO offload support
4  *	Linux INET implementation
5  *
6  *	UDPv4 GSO support
7  */
8 
9 #include <linux/skbuff.h>
10 #include <net/gro.h>
11 #include <net/gso.h>
12 #include <net/udp.h>
13 #include <net/protocol.h>
14 #include <net/inet_common.h>
15 #include <net/udp_tunnel.h>
16 
17 #if IS_ENABLED(CONFIG_NET_UDP_TUNNEL)
18 
19 /*
20  * Dummy GRO tunnel callback, exists mainly to avoid dangling/NULL
21  * values for the udp tunnel static call.
22  */
23 static struct sk_buff *dummy_gro_rcv(struct sock *sk,
24 				     struct list_head *head,
25 				     struct sk_buff *skb)
26 {
27 	NAPI_GRO_CB(skb)->flush = 1;
28 	return NULL;
29 }
30 
31 typedef struct sk_buff *(*udp_tunnel_gro_rcv_t)(struct sock *sk,
32 						struct list_head *head,
33 						struct sk_buff *skb);
34 
35 struct udp_tunnel_type_entry {
36 	udp_tunnel_gro_rcv_t gro_receive;
37 	refcount_t count;
38 };
39 
40 #define UDP_MAX_TUNNEL_TYPES (IS_ENABLED(CONFIG_GENEVE) + \
41 			      IS_ENABLED(CONFIG_VXLAN) * 2 + \
42 			      IS_ENABLED(CONFIG_NET_FOU) * 2 + \
43 			      IS_ENABLED(CONFIG_XFRM) * 2)
44 
45 DEFINE_STATIC_CALL(udp_tunnel_gro_rcv, dummy_gro_rcv);
46 static DEFINE_STATIC_KEY_FALSE(udp_tunnel_static_call);
47 static DEFINE_MUTEX(udp_tunnel_gro_type_lock);
48 static struct udp_tunnel_type_entry udp_tunnel_gro_types[UDP_MAX_TUNNEL_TYPES];
49 static unsigned int udp_tunnel_gro_type_nr;
50 static DEFINE_SPINLOCK(udp_tunnel_gro_lock);
51 
52 void udp_tunnel_update_gro_lookup(struct net *net, struct sock *sk, bool add)
53 {
54 	bool is_ipv6 = sk->sk_family == AF_INET6;
55 	struct udp_sock *tup, *up = udp_sk(sk);
56 	struct udp_tunnel_gro *udp_tunnel_gro;
57 
58 	spin_lock(&udp_tunnel_gro_lock);
59 	udp_tunnel_gro = &net->ipv4.udp_tunnel_gro[is_ipv6];
60 	if (add)
61 		hlist_add_head(&up->tunnel_list, &udp_tunnel_gro->list);
62 	else if (up->tunnel_list.pprev)
63 		hlist_del_init(&up->tunnel_list);
64 
65 	if (udp_tunnel_gro->list.first &&
66 	    !udp_tunnel_gro->list.first->next) {
67 		tup = hlist_entry(udp_tunnel_gro->list.first, struct udp_sock,
68 				  tunnel_list);
69 
70 		rcu_assign_pointer(udp_tunnel_gro->sk, (struct sock *)tup);
71 	} else {
72 		RCU_INIT_POINTER(udp_tunnel_gro->sk, NULL);
73 	}
74 
75 	spin_unlock(&udp_tunnel_gro_lock);
76 }
77 EXPORT_SYMBOL_GPL(udp_tunnel_update_gro_lookup);
78 
79 void udp_tunnel_update_gro_rcv(struct sock *sk, bool add)
80 {
81 	struct udp_tunnel_type_entry *cur = NULL;
82 	struct udp_sock *up = udp_sk(sk);
83 	int i, old_gro_type_nr;
84 
85 	if (!UDP_MAX_TUNNEL_TYPES || !up->gro_receive)
86 		return;
87 
88 	mutex_lock(&udp_tunnel_gro_type_lock);
89 
90 	/* Check if the static call is permanently disabled. */
91 	if (udp_tunnel_gro_type_nr > UDP_MAX_TUNNEL_TYPES)
92 		goto out;
93 
94 	for (i = 0; i < udp_tunnel_gro_type_nr; i++)
95 		if (udp_tunnel_gro_types[i].gro_receive == up->gro_receive)
96 			cur = &udp_tunnel_gro_types[i];
97 
98 	old_gro_type_nr = udp_tunnel_gro_type_nr;
99 	if (add) {
100 		/*
101 		 * Update the matching entry, if found, or add a new one
102 		 * if needed
103 		 */
104 		if (cur) {
105 			refcount_inc(&cur->count);
106 			goto out;
107 		}
108 
109 		if (unlikely(udp_tunnel_gro_type_nr == UDP_MAX_TUNNEL_TYPES)) {
110 			pr_err_once("Too many UDP tunnel types, please increase UDP_MAX_TUNNEL_TYPES\n");
111 			/* Ensure static call will never be enabled */
112 			udp_tunnel_gro_type_nr = UDP_MAX_TUNNEL_TYPES + 1;
113 		} else {
114 			cur = &udp_tunnel_gro_types[udp_tunnel_gro_type_nr++];
115 			refcount_set(&cur->count, 1);
116 			cur->gro_receive = up->gro_receive;
117 		}
118 	} else {
119 		/*
120 		 * The stack cleanups only successfully added tunnel, the
121 		 * lookup on removal should never fail.
122 		 */
123 		if (WARN_ON_ONCE(!cur))
124 			goto out;
125 
126 		if (!refcount_dec_and_test(&cur->count))
127 			goto out;
128 
129 		/* Avoid gaps, so that the enable tunnel has always id 0 */
130 		*cur = udp_tunnel_gro_types[--udp_tunnel_gro_type_nr];
131 	}
132 
133 	if (udp_tunnel_gro_type_nr == 1) {
134 		static_call_update(udp_tunnel_gro_rcv,
135 				   udp_tunnel_gro_types[0].gro_receive);
136 		static_branch_enable(&udp_tunnel_static_call);
137 	} else if (old_gro_type_nr == 1) {
138 		static_branch_disable(&udp_tunnel_static_call);
139 		static_call_update(udp_tunnel_gro_rcv, dummy_gro_rcv);
140 	}
141 
142 out:
143 	mutex_unlock(&udp_tunnel_gro_type_lock);
144 }
145 EXPORT_SYMBOL_GPL(udp_tunnel_update_gro_rcv);
146 
147 static struct sk_buff *udp_tunnel_gro_rcv(struct sock *sk,
148 					  struct list_head *head,
149 					  struct sk_buff *skb)
150 {
151 	if (static_branch_likely(&udp_tunnel_static_call)) {
152 		if (unlikely(gro_recursion_inc_test(skb))) {
153 			NAPI_GRO_CB(skb)->flush |= 1;
154 			return NULL;
155 		}
156 		return static_call(udp_tunnel_gro_rcv)(sk, head, skb);
157 	}
158 	return call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
159 }
160 
161 #else
162 
163 static struct sk_buff *udp_tunnel_gro_rcv(struct sock *sk,
164 					  struct list_head *head,
165 					  struct sk_buff *skb)
166 {
167 	return call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
168 }
169 
170 #endif
171 
172 static struct sk_buff *__skb_udp_tunnel_segment(struct sk_buff *skb,
173 	netdev_features_t features,
174 	struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
175 					     netdev_features_t features),
176 	__be16 new_protocol, bool is_ipv6)
177 {
178 	int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
179 	bool remcsum, need_csum, offload_csum, gso_partial;
180 	struct sk_buff *segs = ERR_PTR(-EINVAL);
181 	struct udphdr *uh = udp_hdr(skb);
182 	u16 mac_offset = skb->mac_header;
183 	__be16 protocol = skb->protocol;
184 	u16 mac_len = skb->mac_len;
185 	int udp_offset, outer_hlen;
186 	__wsum partial;
187 	bool need_ipsec;
188 
189 	if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
190 		goto out;
191 
192 	/* Adjust partial header checksum to negate old length.
193 	 * We cannot rely on the value contained in uh->len as it is
194 	 * possible that the actual value exceeds the boundaries of the
195 	 * 16 bit length field due to the header being added outside of an
196 	 * IP or IPv6 frame that was already limited to 64K - 1.
197 	 */
198 	if (skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)
199 		partial = (__force __wsum)uh->len;
200 	else
201 		partial = (__force __wsum)htonl(skb->len);
202 	partial = csum_sub(csum_unfold(uh->check), partial);
203 
204 	/* setup inner skb. */
205 	skb->encapsulation = 0;
206 	SKB_GSO_CB(skb)->encap_level = 0;
207 	__skb_pull(skb, tnl_hlen);
208 	skb_reset_mac_header(skb);
209 	skb_set_network_header(skb, skb_inner_network_offset(skb));
210 	skb_set_transport_header(skb, skb_inner_transport_offset(skb));
211 	skb->mac_len = skb_inner_network_offset(skb);
212 	skb->protocol = new_protocol;
213 
214 	need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
215 	skb->encap_hdr_csum = need_csum;
216 
217 	remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
218 	skb->remcsum_offload = remcsum;
219 
220 	need_ipsec = (skb_dst(skb) && dst_xfrm(skb_dst(skb))) || skb_sec_path(skb);
221 	/* Try to offload checksum if possible */
222 	offload_csum = !!(need_csum &&
223 			  !need_ipsec &&
224 			  (skb->dev->features &
225 			   (is_ipv6 ? (NETIF_F_HW_CSUM | NETIF_F_IPV6_CSUM) :
226 				      (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM))));
227 
228 	features &= skb->dev->hw_enc_features;
229 	if (need_csum)
230 		features &= ~NETIF_F_SCTP_CRC;
231 
232 	/* The only checksum offload we care about from here on out is the
233 	 * outer one so strip the existing checksum feature flags and
234 	 * instead set the flag based on our outer checksum offload value.
235 	 */
236 	if (remcsum) {
237 		features &= ~NETIF_F_CSUM_MASK;
238 		if (!need_csum || offload_csum)
239 			features |= NETIF_F_HW_CSUM;
240 	}
241 
242 	/* segment inner packet. */
243 	segs = gso_inner_segment(skb, features);
244 	if (IS_ERR_OR_NULL(segs)) {
245 		skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
246 				     mac_len);
247 		goto out;
248 	}
249 
250 	gso_partial = !!(skb_shinfo(segs)->gso_type & SKB_GSO_PARTIAL);
251 
252 	outer_hlen = skb_tnl_header_len(skb);
253 	udp_offset = outer_hlen - tnl_hlen;
254 	skb = segs;
255 	do {
256 		unsigned int len;
257 
258 		if (remcsum)
259 			skb->ip_summed = CHECKSUM_NONE;
260 
261 		/* Set up inner headers if we are offloading inner checksum */
262 		if (skb->ip_summed == CHECKSUM_PARTIAL) {
263 			skb_reset_inner_headers(skb);
264 			skb->encapsulation = 1;
265 		}
266 
267 		skb->mac_len = mac_len;
268 		skb->protocol = protocol;
269 
270 		__skb_push(skb, outer_hlen);
271 		skb_reset_mac_header(skb);
272 		skb_set_network_header(skb, mac_len);
273 		skb_set_transport_header(skb, udp_offset);
274 		len = skb->len - udp_offset;
275 		uh = udp_hdr(skb);
276 
277 		/* If we are only performing partial GSO the inner header
278 		 * will be using a length value equal to only one MSS sized
279 		 * segment instead of the entire frame.
280 		 */
281 		if (gso_partial && skb_is_gso(skb)) {
282 			udp_set_len_short(uh, skb_shinfo(skb)->gso_size +
283 					  SKB_GSO_CB(skb)->data_offset +
284 					  skb->head - (unsigned char *)uh);
285 		} else {
286 			udp_set_len_short(uh, len);
287 		}
288 
289 		if (!need_csum)
290 			continue;
291 
292 		uh->check = ~csum_fold(csum_add(partial,
293 				       (__force __wsum)htonl(len)));
294 
295 		if (skb->encapsulation || !offload_csum) {
296 			uh->check = gso_make_checksum(skb, ~uh->check);
297 			if (uh->check == 0)
298 				uh->check = CSUM_MANGLED_0;
299 		} else {
300 			skb->ip_summed = CHECKSUM_PARTIAL;
301 			skb->csum_start = skb_transport_header(skb) - skb->head;
302 			skb->csum_offset = offsetof(struct udphdr, check);
303 		}
304 	} while ((skb = skb->next));
305 out:
306 	return segs;
307 }
308 
309 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
310 				       netdev_features_t features,
311 				       bool is_ipv6)
312 {
313 	const struct net_offload __rcu **offloads;
314 	__be16 protocol = skb->protocol;
315 	const struct net_offload *ops;
316 	struct sk_buff *segs = ERR_PTR(-EINVAL);
317 	struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
318 					     netdev_features_t features);
319 
320 	rcu_read_lock();
321 
322 	switch (skb->inner_protocol_type) {
323 	case ENCAP_TYPE_ETHER:
324 		protocol = skb->inner_protocol;
325 		gso_inner_segment = skb_mac_gso_segment;
326 		break;
327 	case ENCAP_TYPE_IPPROTO:
328 		offloads = is_ipv6 ? inet6_offloads : inet_offloads;
329 		ops = rcu_dereference(offloads[skb->inner_ipproto]);
330 		if (!ops || !ops->callbacks.gso_segment)
331 			goto out_unlock;
332 		gso_inner_segment = ops->callbacks.gso_segment;
333 		break;
334 	default:
335 		goto out_unlock;
336 	}
337 
338 	segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
339 					protocol, is_ipv6);
340 
341 out_unlock:
342 	rcu_read_unlock();
343 
344 	return segs;
345 }
346 
347 static void __udpv4_gso_segment_csum(struct sk_buff *seg,
348 				     __be32 *oldip, __be32 *newip,
349 				     __be16 *oldport, __be16 *newport)
350 {
351 	struct udphdr *uh;
352 	struct iphdr *iph;
353 
354 	if (*oldip == *newip && *oldport == *newport)
355 		return;
356 
357 	uh = udp_hdr(seg);
358 	iph = ip_hdr(seg);
359 
360 	if (uh->check) {
361 		inet_proto_csum_replace4(&uh->check, seg, *oldip, *newip,
362 					 true);
363 		inet_proto_csum_replace2(&uh->check, seg, *oldport, *newport,
364 					 false);
365 		if (!uh->check)
366 			uh->check = CSUM_MANGLED_0;
367 	}
368 	*oldport = *newport;
369 
370 	csum_replace4(&iph->check, *oldip, *newip);
371 	*oldip = *newip;
372 }
373 
374 static struct sk_buff *__udpv4_gso_segment_list_csum(struct sk_buff *segs)
375 {
376 	struct sk_buff *seg;
377 	struct udphdr *uh, *uh2;
378 	struct iphdr *iph, *iph2;
379 
380 	seg = segs;
381 	uh = udp_hdr(seg);
382 	iph = ip_hdr(seg);
383 
384 	if ((udp_hdr(seg)->dest == udp_hdr(seg->next)->dest) &&
385 	    (udp_hdr(seg)->source == udp_hdr(seg->next)->source) &&
386 	    (ip_hdr(seg)->daddr == ip_hdr(seg->next)->daddr) &&
387 	    (ip_hdr(seg)->saddr == ip_hdr(seg->next)->saddr))
388 		return segs;
389 
390 	while ((seg = seg->next)) {
391 		uh2 = udp_hdr(seg);
392 		iph2 = ip_hdr(seg);
393 
394 		__udpv4_gso_segment_csum(seg,
395 					 &iph2->saddr, &iph->saddr,
396 					 &uh2->source, &uh->source);
397 		__udpv4_gso_segment_csum(seg,
398 					 &iph2->daddr, &iph->daddr,
399 					 &uh2->dest, &uh->dest);
400 	}
401 
402 	return segs;
403 }
404 
405 static void __udpv6_gso_segment_csum(struct sk_buff *seg,
406 				     struct in6_addr *oldip,
407 				     const struct in6_addr *newip,
408 				     __be16 *oldport, __be16 newport)
409 {
410 	struct udphdr *uh = udp_hdr(seg);
411 
412 	if (ipv6_addr_equal(oldip, newip) && *oldport == newport)
413 		return;
414 
415 	if (uh->check) {
416 		inet_proto_csum_replace16(&uh->check, seg, oldip->s6_addr32,
417 					  newip->s6_addr32, true);
418 
419 		inet_proto_csum_replace2(&uh->check, seg, *oldport, newport,
420 					 false);
421 		if (!uh->check)
422 			uh->check = CSUM_MANGLED_0;
423 	}
424 
425 	*oldip = *newip;
426 	*oldport = newport;
427 }
428 
429 static struct sk_buff *__udpv6_gso_segment_list_csum(struct sk_buff *segs)
430 {
431 	const struct ipv6hdr *iph;
432 	const struct udphdr *uh;
433 	struct ipv6hdr *iph2;
434 	struct sk_buff *seg;
435 	struct udphdr *uh2;
436 
437 	seg = segs;
438 	uh = udp_hdr(seg);
439 	iph = ipv6_hdr(seg);
440 	uh2 = udp_hdr(seg->next);
441 	iph2 = ipv6_hdr(seg->next);
442 
443 	if (!(*(const u32 *)&uh->source ^ *(const u32 *)&uh2->source) &&
444 	    ipv6_addr_equal(&iph->saddr, &iph2->saddr) &&
445 	    ipv6_addr_equal(&iph->daddr, &iph2->daddr))
446 		return segs;
447 
448 	while ((seg = seg->next)) {
449 		uh2 = udp_hdr(seg);
450 		iph2 = ipv6_hdr(seg);
451 
452 		__udpv6_gso_segment_csum(seg, &iph2->saddr, &iph->saddr,
453 					 &uh2->source, uh->source);
454 		__udpv6_gso_segment_csum(seg, &iph2->daddr, &iph->daddr,
455 					 &uh2->dest, uh->dest);
456 	}
457 
458 	return segs;
459 }
460 
461 static struct sk_buff *__udp_gso_segment_list(struct sk_buff *skb,
462 					      netdev_features_t features,
463 					      bool is_ipv6)
464 {
465 	unsigned int mss = skb_shinfo(skb)->gso_size;
466 
467 	skb = skb_segment_list(skb, features, skb_mac_header_len(skb));
468 	if (IS_ERR(skb))
469 		return skb;
470 
471 	udp_set_len_short(udp_hdr(skb), sizeof(struct udphdr) + mss);
472 
473 	if (is_ipv6)
474 		return __udpv6_gso_segment_list_csum(skb);
475 	else
476 		return __udpv4_gso_segment_list_csum(skb);
477 }
478 
479 struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
480 				  netdev_features_t features, bool is_ipv6)
481 {
482 	struct sock *sk = gso_skb->sk;
483 	unsigned int sum_truesize = 0;
484 	struct sk_buff *segs, *seg;
485 	struct udphdr *uh;
486 	unsigned int mss;
487 	bool copy_dtor;
488 	__sum16 check;
489 	int ret = 0;
490 	u16 newlen;
491 
492 	mss = skb_shinfo(gso_skb)->gso_size;
493 	if (gso_skb->len <= sizeof(*uh) + mss)
494 		return ERR_PTR(-EINVAL);
495 
496 	if (unlikely(skb_checksum_start(gso_skb) !=
497 		     skb_transport_header(gso_skb) &&
498 		     !(skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST)))
499 		return ERR_PTR(-EINVAL);
500 
501 	/* We don't know if egress device can segment and checksum the packet
502 	 * when IPv6 extension headers are present. Fall back to software GSO.
503 	 */
504 	if (gso_skb->ip_summed != CHECKSUM_PARTIAL)
505 		features &= ~(NETIF_F_GSO_UDP_L4 | NETIF_F_CSUM_MASK);
506 
507 	if (skb_gso_ok(gso_skb, features | NETIF_F_GSO_ROBUST)) {
508 		/* Packet is from an untrusted source, reset gso_segs. */
509 		skb_shinfo(gso_skb)->gso_segs = DIV_ROUND_UP(gso_skb->len - sizeof(*uh),
510 							     mss);
511 		return NULL;
512 	}
513 
514 	if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST) {
515 		 /* Detect modified geometry and pass those to skb_segment. */
516 		if ((skb_pagelen(gso_skb) - sizeof(*uh) == skb_shinfo(gso_skb)->gso_size) &&
517 		    !(skb_shinfo(gso_skb)->gso_type & SKB_GSO_DODGY))
518 			return __udp_gso_segment_list(gso_skb, features, is_ipv6);
519 
520 		ret = __skb_linearize(gso_skb);
521 		if (ret)
522 			return ERR_PTR(ret);
523 
524 		 /* Setup csum, as fraglist skips this in udp4_gro_receive. */
525 		gso_skb->csum_start = skb_transport_header(gso_skb) - gso_skb->head;
526 		gso_skb->csum_offset = offsetof(struct udphdr, check);
527 		gso_skb->ip_summed = CHECKSUM_PARTIAL;
528 
529 		uh = udp_hdr(gso_skb);
530 		if (is_ipv6)
531 			uh->check = ~udp_v6_check(gso_skb->len,
532 						  &ipv6_hdr(gso_skb)->saddr,
533 						  &ipv6_hdr(gso_skb)->daddr, 0);
534 		else
535 			uh->check = ~udp_v4_check(gso_skb->len,
536 						  ip_hdr(gso_skb)->saddr,
537 						  ip_hdr(gso_skb)->daddr, 0);
538 	}
539 
540 	skb_pull(gso_skb, sizeof(*uh));
541 
542 	/* clear destructor to avoid skb_segment assigning it to tail */
543 	copy_dtor = gso_skb->destructor == sock_wfree;
544 	if (copy_dtor) {
545 		gso_skb->destructor = NULL;
546 		gso_skb->sk = NULL;
547 	}
548 
549 	segs = skb_segment(gso_skb, features);
550 	if (IS_ERR_OR_NULL(segs)) {
551 		if (copy_dtor) {
552 			gso_skb->destructor = sock_wfree;
553 			gso_skb->sk = sk;
554 		}
555 		return segs;
556 	}
557 
558 	seg = segs;
559 	uh = udp_hdr(seg);
560 
561 	/* preserve TX timestamp flags and TS key for first segment */
562 	skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey;
563 	skb_shinfo(seg)->tx_flags |=
564 			(skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP);
565 
566 	/* compute checksum adjustment based on old length versus new */
567 	newlen = sizeof(*uh) + mss;
568 	check = csum16_add(csum16_sub(uh->check, uh->len), htons(newlen));
569 
570 	for (;;) {
571 		if (copy_dtor) {
572 			seg->destructor = sock_wfree;
573 			seg->sk = sk;
574 			sum_truesize += seg->truesize;
575 		}
576 
577 		if (!seg->next)
578 			break;
579 
580 		udp_set_len_short(uh, newlen);
581 		uh->check = check;
582 
583 		if (seg->ip_summed == CHECKSUM_PARTIAL)
584 			gso_reset_checksum(seg, ~check);
585 		else
586 			uh->check = gso_make_checksum(seg, ~check) ? :
587 				    CSUM_MANGLED_0;
588 
589 		seg = seg->next;
590 		uh = udp_hdr(seg);
591 	}
592 
593 	/* Unless skb fits perfectly as GSO_PARTIAL, the trailing
594 	 * segment may not be full MSS, account for that in the checksum
595 	 */
596 	if (!skb_is_gso(seg))
597 		newlen = skb_tail_pointer(seg) - skb_transport_header(seg) + seg->data_len;
598 	check = csum16_add(csum16_sub(uh->check, uh->len), htons(newlen));
599 
600 	udp_set_len_short(uh, newlen);
601 	uh->check = check;
602 
603 	if (seg->ip_summed == CHECKSUM_PARTIAL)
604 		gso_reset_checksum(seg, ~check);
605 	else
606 		uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0;
607 
608 	/* On the TX path, CHECKSUM_NONE and CHECKSUM_UNNECESSARY have the same
609 	 * meaning. However, check for bad offloads in the GSO stack expects the
610 	 * latter, if the checksum was calculated in software. To vouch for the
611 	 * segment skbs we actually need to set it on the gso_skb.
612 	 */
613 	if (gso_skb->ip_summed == CHECKSUM_NONE)
614 		gso_skb->ip_summed = CHECKSUM_UNNECESSARY;
615 
616 	/* update refcount for the packet */
617 	if (copy_dtor) {
618 		int delta = sum_truesize - gso_skb->truesize;
619 
620 		/* In some pathological cases, delta can be negative.
621 		 * We need to either use refcount_add() or refcount_sub_and_test()
622 		 */
623 		if (likely(delta >= 0))
624 			refcount_add(delta, &sk->sk_wmem_alloc);
625 		else
626 			WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));
627 	}
628 	return segs;
629 }
630 
631 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
632 					 netdev_features_t features)
633 {
634 	struct sk_buff *segs = ERR_PTR(-EINVAL);
635 	unsigned int mss;
636 	__wsum csum;
637 	struct udphdr *uh;
638 	struct iphdr *iph;
639 
640 	if (skb->encapsulation &&
641 	    (skb_shinfo(skb)->gso_type &
642 	     (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
643 		segs = skb_udp_tunnel_segment(skb, features, false);
644 		goto out;
645 	}
646 
647 	if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4)))
648 		goto out;
649 
650 	if (!pskb_may_pull(skb, sizeof(struct udphdr)))
651 		goto out;
652 
653 	if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
654 		return __udp_gso_segment(skb, features, false);
655 
656 	mss = skb_shinfo(skb)->gso_size;
657 	if (unlikely(skb->len <= mss))
658 		goto out;
659 
660 	/* Do software UFO. Complete and fill in the UDP checksum as
661 	 * HW cannot do checksum of UDP packets sent as multiple
662 	 * IP fragments.
663 	 */
664 
665 	uh = udp_hdr(skb);
666 	iph = ip_hdr(skb);
667 
668 	uh->check = 0;
669 	csum = skb_checksum(skb, 0, skb->len, 0);
670 	uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
671 	if (uh->check == 0)
672 		uh->check = CSUM_MANGLED_0;
673 
674 	skb->ip_summed = CHECKSUM_UNNECESSARY;
675 
676 	/* If there is no outer header we can fake a checksum offload
677 	 * due to the fact that we have already done the checksum in
678 	 * software prior to segmenting the frame.
679 	 */
680 	if (!skb->encap_hdr_csum)
681 		features |= NETIF_F_HW_CSUM;
682 
683 	/* Fragment the skb. IP headers of the fragments are updated in
684 	 * inet_gso_segment()
685 	 */
686 	segs = skb_segment(skb, features);
687 out:
688 	return segs;
689 }
690 
691 
692 #define UDP_GRO_CNT_MAX 64
693 static struct sk_buff *udp_gro_receive_segment(struct list_head *head,
694 					       struct sk_buff *skb)
695 {
696 	struct udphdr *uh = udp_gro_udphdr(skb);
697 	struct sk_buff *pp = NULL;
698 	struct udphdr *uh2;
699 	struct sk_buff *p;
700 	unsigned int ulen;
701 	int ret = 0;
702 	int flush;
703 
704 	/* requires non zero csum, for symmetry with GSO */
705 	if (!uh->check) {
706 		NAPI_GRO_CB(skb)->flush = 1;
707 		return NULL;
708 	}
709 
710 	ulen = udp_get_len_short(uh);
711 
712 	/* pull encapsulating udp header */
713 	skb_gro_pull(skb, sizeof(struct udphdr));
714 
715 	list_for_each_entry(p, head, list) {
716 		if (!NAPI_GRO_CB(p)->same_flow)
717 			continue;
718 
719 		uh2 = udp_hdr(p);
720 
721 		/* Match ports only, as csum is always non zero */
722 		if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) {
723 			NAPI_GRO_CB(p)->same_flow = 0;
724 			continue;
725 		}
726 
727 		if (NAPI_GRO_CB(skb)->is_flist != NAPI_GRO_CB(p)->is_flist) {
728 			NAPI_GRO_CB(skb)->flush = 1;
729 			return p;
730 		}
731 
732 		flush = gro_receive_network_flush(uh, uh2, p);
733 
734 		/* Terminate the flow on len mismatch or if it grow "too much".
735 		 * Under small packet flood GRO count could elsewhere grow a lot
736 		 * leading to excessive truesize values.
737 		 * On len mismatch merge the first packet shorter than gso_size,
738 		 * otherwise complete the GRO packet.
739 		 */
740 		if (ulen > udp_get_len_short(uh2) || flush) {
741 			pp = p;
742 		} else {
743 			if (NAPI_GRO_CB(skb)->is_flist) {
744 				if (!pskb_may_pull(skb, skb_gro_offset(skb))) {
745 					NAPI_GRO_CB(skb)->flush = 1;
746 					return NULL;
747 				}
748 				if ((skb->ip_summed != p->ip_summed) ||
749 				    (skb->csum_level != p->csum_level)) {
750 					NAPI_GRO_CB(skb)->flush = 1;
751 					return NULL;
752 				}
753 				skb_set_network_header(skb, skb_gro_receive_network_offset(skb));
754 				ret = skb_gro_receive_list(p, skb);
755 			} else {
756 				skb_gro_postpull_rcsum(skb, uh,
757 						       sizeof(struct udphdr));
758 
759 				ret = skb_gro_receive(p, skb);
760 			}
761 		}
762 
763 		if (ret || ulen != udp_get_len_short(uh2) ||
764 		    NAPI_GRO_CB(p)->count >= UDP_GRO_CNT_MAX)
765 			pp = p;
766 
767 		return pp;
768 	}
769 
770 	/* mismatch, but we never need to flush */
771 	return NULL;
772 }
773 
774 struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb,
775 				struct udphdr *uh, struct sock *sk)
776 {
777 	struct sk_buff *pp = NULL;
778 	struct sk_buff *p;
779 	struct udphdr *uh2;
780 	unsigned int off = skb_gro_offset(skb);
781 	unsigned int ulen;
782 	int flush = 1;
783 
784 	/* Do not deal with padded or malicious packets, sorry! */
785 	ulen = udp_get_len_short(uh);
786 	if (ulen <= sizeof(*uh) || ulen != skb_gro_len(skb))
787 		goto out;
788 
789 	/* We can do L4 aggregation only if the packet can't land in a tunnel
790 	 * otherwise we could corrupt the inner stream. Detecting such packets
791 	 * cannot be foolproof and the aggregation might still happen in some
792 	 * cases. Such packets should be caught in udp_unexpected_gso later.
793 	 */
794 	NAPI_GRO_CB(skb)->is_flist = 0;
795 	if (!sk || !udp_sk(sk)->gro_receive) {
796 		/* If the packet was locally encapsulated in a UDP tunnel that
797 		 * wasn't detected above, do not GRO.
798 		 */
799 		if (skb->encapsulation)
800 			goto out;
801 
802 		if (skb->dev->features & NETIF_F_GRO_FRAGLIST)
803 			NAPI_GRO_CB(skb)->is_flist = sk ? !udp_test_bit(GRO_ENABLED, sk) : 1;
804 
805 		if ((!sk && (skb->dev->features & NETIF_F_GRO_UDP_FWD)) ||
806 		    (sk && udp_test_bit(GRO_ENABLED, sk)) || NAPI_GRO_CB(skb)->is_flist)
807 			return call_gro_receive(udp_gro_receive_segment, head, skb);
808 
809 		/* no GRO, be sure flush the current packet */
810 		goto out;
811 	}
812 
813 	if (NAPI_GRO_CB(skb)->encap_mark ||
814 	    (uh->check && skb->ip_summed != CHECKSUM_PARTIAL &&
815 	     NAPI_GRO_CB(skb)->csum_cnt == 0 &&
816 	     !NAPI_GRO_CB(skb)->csum_valid))
817 		goto out;
818 
819 	/* mark that this skb passed once through the tunnel gro layer */
820 	NAPI_GRO_CB(skb)->encap_mark = 1;
821 
822 	flush = 0;
823 
824 	list_for_each_entry(p, head, list) {
825 		if (!NAPI_GRO_CB(p)->same_flow)
826 			continue;
827 
828 		uh2 = (struct udphdr   *)(p->data + off);
829 
830 		/* Match ports and either checksums are either both zero
831 		 * or nonzero.
832 		 */
833 		if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
834 		    (!uh->check ^ !uh2->check)) {
835 			NAPI_GRO_CB(p)->same_flow = 0;
836 			continue;
837 		}
838 	}
839 
840 	skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
841 	skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
842 	pp = udp_tunnel_gro_rcv(sk, head, skb);
843 
844 out:
845 	skb_gro_flush_final(skb, pp, flush);
846 	return pp;
847 }
848 
849 static struct sock *udp4_gro_lookup_skb(struct sk_buff *skb, __be16 sport,
850 					__be16 dport)
851 {
852 	const struct iphdr *iph = skb_gro_network_header(skb);
853 	struct net *net = dev_net_rcu(skb->dev);
854 	struct sock *sk;
855 	int iif, sdif;
856 
857 	sk = udp_tunnel_sk(net, false);
858 	if (sk && dport == htons(sk->sk_num))
859 		return sk;
860 
861 	inet_get_iif_sdif(skb, &iif, &sdif);
862 
863 	return __udp4_lib_lookup(net, iph->saddr, sport,
864 				 iph->daddr, dport, iif, sdif, NULL);
865 }
866 
867 INDIRECT_CALLABLE_SCOPE
868 struct sk_buff *udp4_gro_receive(struct list_head *head, struct sk_buff *skb)
869 {
870 	struct udphdr *uh = udp_gro_udphdr(skb);
871 	struct sock *sk = NULL;
872 	struct sk_buff *pp;
873 
874 	if (unlikely(!uh))
875 		goto flush;
876 
877 	/* Don't bother verifying checksum if we're going to flush anyway. */
878 	if (NAPI_GRO_CB(skb)->flush)
879 		goto skip;
880 
881 	if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
882 						 inet_gro_compute_pseudo))
883 		goto flush;
884 	else if (uh->check)
885 		skb_gro_checksum_try_convert(skb, IPPROTO_UDP,
886 					     inet_gro_compute_pseudo);
887 skip:
888 	if (static_branch_unlikely(&udp_encap_needed_key))
889 		sk = udp4_gro_lookup_skb(skb, uh->source, uh->dest);
890 
891 	pp = udp_gro_receive(head, skb, uh, sk);
892 	return pp;
893 
894 flush:
895 	NAPI_GRO_CB(skb)->flush = 1;
896 	return NULL;
897 }
898 
899 static int udp_gro_complete_segment(struct sk_buff *skb)
900 {
901 	struct udphdr *uh = udp_hdr(skb);
902 
903 	skb->csum_start = (unsigned char *)uh - skb->head;
904 	skb->csum_offset = offsetof(struct udphdr, check);
905 	skb->ip_summed = CHECKSUM_PARTIAL;
906 
907 	skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
908 	skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_L4;
909 
910 	if (skb->encapsulation)
911 		skb->inner_transport_header = skb->transport_header;
912 
913 	return 0;
914 }
915 
916 int udp_gro_complete(struct sk_buff *skb, int nhoff,
917 		     udp_lookup_t lookup)
918 {
919 	struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
920 	unsigned int newlen = skb->len - nhoff;
921 	struct sock *sk;
922 	int err;
923 
924 	udp_set_len(uh, newlen);
925 
926 	sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb,
927 				udp4_lib_lookup_skb, skb, uh->source, uh->dest);
928 	if (sk && udp_sk(sk)->gro_complete) {
929 		skb_shinfo(skb)->gso_type = uh->check ? SKB_GSO_UDP_TUNNEL_CSUM
930 					: SKB_GSO_UDP_TUNNEL;
931 
932 		/* clear the encap mark, so that inner frag_list gro_complete
933 		 * can take place
934 		 */
935 		NAPI_GRO_CB(skb)->encap_mark = 0;
936 
937 		/* Set encapsulation before calling into inner gro_complete()
938 		 * functions to make them set up the inner offsets.
939 		 */
940 		skb->encapsulation = 1;
941 		err = udp_sk(sk)->gro_complete(sk, skb,
942 				nhoff + sizeof(struct udphdr));
943 	} else {
944 		err = udp_gro_complete_segment(skb);
945 	}
946 
947 	if (skb->remcsum_offload)
948 		skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
949 
950 	return err;
951 }
952 
953 INDIRECT_CALLABLE_SCOPE int udp4_gro_complete(struct sk_buff *skb, int nhoff)
954 {
955 	const u16 offset = NAPI_GRO_CB(skb)->network_offsets[skb->encapsulation];
956 	const struct iphdr *iph = (struct iphdr *)(skb->data + offset);
957 	struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
958 
959 	/* do fraglist only if there is no outer UDP encap (or we already processed it) */
960 	if (NAPI_GRO_CB(skb)->is_flist && !NAPI_GRO_CB(skb)->encap_mark) {
961 		udp_set_len(uh, skb->len - nhoff);
962 
963 		skb_shinfo(skb)->gso_type |= (SKB_GSO_FRAGLIST|SKB_GSO_UDP_L4);
964 		skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
965 
966 		__skb_incr_checksum_unnecessary(skb);
967 
968 		return 0;
969 	}
970 
971 	if (uh->check)
972 		uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
973 					  iph->daddr, 0);
974 
975 	return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
976 }
977 
978 int __init udpv4_offload_init(void)
979 {
980 	net_hotdata.udpv4_offload = (struct net_offload) {
981 		.callbacks = {
982 			.gso_segment = udp4_ufo_fragment,
983 			.gro_receive  =	udp4_gro_receive,
984 			.gro_complete =	udp4_gro_complete,
985 		},
986 	};
987 
988 	return inet_add_offload(&net_hotdata.udpv4_offload, IPPROTO_UDP);
989 }
990