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