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 */
dummy_gro_rcv(struct sock * sk,struct list_head * head,struct sk_buff * skb)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 struct 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
udp_tunnel_update_gro_lookup(struct net * net,struct sock * sk,bool add)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
udp_tunnel_update_gro_rcv(struct sock * sk,bool add)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
udp_tunnel_gro_init(void)147 static void udp_tunnel_gro_init(void)
148 {
149 mutex_init(&udp_tunnel_gro_type_lock);
150 }
151
udp_tunnel_gro_rcv(struct sock * sk,struct list_head * head,struct sk_buff * skb)152 static struct sk_buff *udp_tunnel_gro_rcv(struct sock *sk,
153 struct list_head *head,
154 struct sk_buff *skb)
155 {
156 if (static_branch_likely(&udp_tunnel_static_call)) {
157 if (unlikely(gro_recursion_inc_test(skb))) {
158 NAPI_GRO_CB(skb)->flush |= 1;
159 return NULL;
160 }
161 return static_call(udp_tunnel_gro_rcv)(sk, head, skb);
162 }
163 return call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
164 }
165
166 #else
167
udp_tunnel_gro_init(void)168 static void udp_tunnel_gro_init(void) {}
169
udp_tunnel_gro_rcv(struct sock * sk,struct list_head * head,struct sk_buff * skb)170 static struct sk_buff *udp_tunnel_gro_rcv(struct sock *sk,
171 struct list_head *head,
172 struct sk_buff *skb)
173 {
174 return call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
175 }
176
177 #endif
178
__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)179 static struct sk_buff *__skb_udp_tunnel_segment(struct sk_buff *skb,
180 netdev_features_t features,
181 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
182 netdev_features_t features),
183 __be16 new_protocol, bool is_ipv6)
184 {
185 int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
186 bool remcsum, need_csum, offload_csum, gso_partial;
187 struct sk_buff *segs = ERR_PTR(-EINVAL);
188 struct udphdr *uh = udp_hdr(skb);
189 u16 mac_offset = skb->mac_header;
190 __be16 protocol = skb->protocol;
191 u16 mac_len = skb->mac_len;
192 int udp_offset, outer_hlen;
193 __wsum partial;
194 bool need_ipsec;
195
196 if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
197 goto out;
198
199 /* Adjust partial header checksum to negate old length.
200 * We cannot rely on the value contained in uh->len as it is
201 * possible that the actual value exceeds the boundaries of the
202 * 16 bit length field due to the header being added outside of an
203 * IP or IPv6 frame that was already limited to 64K - 1.
204 */
205 if (skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)
206 partial = (__force __wsum)uh->len;
207 else
208 partial = (__force __wsum)htonl(skb->len);
209 partial = csum_sub(csum_unfold(uh->check), partial);
210
211 /* setup inner skb. */
212 skb->encapsulation = 0;
213 SKB_GSO_CB(skb)->encap_level = 0;
214 __skb_pull(skb, tnl_hlen);
215 skb_reset_mac_header(skb);
216 skb_set_network_header(skb, skb_inner_network_offset(skb));
217 skb_set_transport_header(skb, skb_inner_transport_offset(skb));
218 skb->mac_len = skb_inner_network_offset(skb);
219 skb->protocol = new_protocol;
220
221 need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
222 skb->encap_hdr_csum = need_csum;
223
224 remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
225 skb->remcsum_offload = remcsum;
226
227 need_ipsec = skb_dst(skb) && dst_xfrm(skb_dst(skb));
228 /* Try to offload checksum if possible */
229 offload_csum = !!(need_csum &&
230 !need_ipsec &&
231 (skb->dev->features &
232 (is_ipv6 ? (NETIF_F_HW_CSUM | NETIF_F_IPV6_CSUM) :
233 (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM))));
234
235 features &= skb->dev->hw_enc_features;
236 if (need_csum)
237 features &= ~NETIF_F_SCTP_CRC;
238
239 /* The only checksum offload we care about from here on out is the
240 * outer one so strip the existing checksum feature flags and
241 * instead set the flag based on our outer checksum offload value.
242 */
243 if (remcsum) {
244 features &= ~NETIF_F_CSUM_MASK;
245 if (!need_csum || offload_csum)
246 features |= NETIF_F_HW_CSUM;
247 }
248
249 /* segment inner packet. */
250 segs = gso_inner_segment(skb, features);
251 if (IS_ERR_OR_NULL(segs)) {
252 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
253 mac_len);
254 goto out;
255 }
256
257 gso_partial = !!(skb_shinfo(segs)->gso_type & SKB_GSO_PARTIAL);
258
259 outer_hlen = skb_tnl_header_len(skb);
260 udp_offset = outer_hlen - tnl_hlen;
261 skb = segs;
262 do {
263 unsigned int len;
264
265 if (remcsum)
266 skb->ip_summed = CHECKSUM_NONE;
267
268 /* Set up inner headers if we are offloading inner checksum */
269 if (skb->ip_summed == CHECKSUM_PARTIAL) {
270 skb_reset_inner_headers(skb);
271 skb->encapsulation = 1;
272 }
273
274 skb->mac_len = mac_len;
275 skb->protocol = protocol;
276
277 __skb_push(skb, outer_hlen);
278 skb_reset_mac_header(skb);
279 skb_set_network_header(skb, mac_len);
280 skb_set_transport_header(skb, udp_offset);
281 len = skb->len - udp_offset;
282 uh = udp_hdr(skb);
283
284 /* If we are only performing partial GSO the inner header
285 * will be using a length value equal to only one MSS sized
286 * segment instead of the entire frame.
287 */
288 if (gso_partial && skb_is_gso(skb)) {
289 uh->len = htons(skb_shinfo(skb)->gso_size +
290 SKB_GSO_CB(skb)->data_offset +
291 skb->head - (unsigned char *)uh);
292 } else {
293 uh->len = htons(len);
294 }
295
296 if (!need_csum)
297 continue;
298
299 uh->check = ~csum_fold(csum_add(partial,
300 (__force __wsum)htonl(len)));
301
302 if (skb->encapsulation || !offload_csum) {
303 uh->check = gso_make_checksum(skb, ~uh->check);
304 if (uh->check == 0)
305 uh->check = CSUM_MANGLED_0;
306 } else {
307 skb->ip_summed = CHECKSUM_PARTIAL;
308 skb->csum_start = skb_transport_header(skb) - skb->head;
309 skb->csum_offset = offsetof(struct udphdr, check);
310 }
311 } while ((skb = skb->next));
312 out:
313 return segs;
314 }
315
skb_udp_tunnel_segment(struct sk_buff * skb,netdev_features_t features,bool is_ipv6)316 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
317 netdev_features_t features,
318 bool is_ipv6)
319 {
320 const struct net_offload __rcu **offloads;
321 __be16 protocol = skb->protocol;
322 const struct net_offload *ops;
323 struct sk_buff *segs = ERR_PTR(-EINVAL);
324 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
325 netdev_features_t features);
326
327 rcu_read_lock();
328
329 switch (skb->inner_protocol_type) {
330 case ENCAP_TYPE_ETHER:
331 protocol = skb->inner_protocol;
332 gso_inner_segment = skb_mac_gso_segment;
333 break;
334 case ENCAP_TYPE_IPPROTO:
335 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
336 ops = rcu_dereference(offloads[skb->inner_ipproto]);
337 if (!ops || !ops->callbacks.gso_segment)
338 goto out_unlock;
339 gso_inner_segment = ops->callbacks.gso_segment;
340 break;
341 default:
342 goto out_unlock;
343 }
344
345 segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
346 protocol, is_ipv6);
347
348 out_unlock:
349 rcu_read_unlock();
350
351 return segs;
352 }
353 EXPORT_SYMBOL(skb_udp_tunnel_segment);
354
__udpv4_gso_segment_csum(struct sk_buff * seg,__be32 * oldip,__be32 * newip,__be16 * oldport,__be16 * newport)355 static void __udpv4_gso_segment_csum(struct sk_buff *seg,
356 __be32 *oldip, __be32 *newip,
357 __be16 *oldport, __be16 *newport)
358 {
359 struct udphdr *uh;
360 struct iphdr *iph;
361
362 if (*oldip == *newip && *oldport == *newport)
363 return;
364
365 uh = udp_hdr(seg);
366 iph = ip_hdr(seg);
367
368 if (uh->check) {
369 inet_proto_csum_replace4(&uh->check, seg, *oldip, *newip,
370 true);
371 inet_proto_csum_replace2(&uh->check, seg, *oldport, *newport,
372 false);
373 if (!uh->check)
374 uh->check = CSUM_MANGLED_0;
375 }
376 *oldport = *newport;
377
378 csum_replace4(&iph->check, *oldip, *newip);
379 *oldip = *newip;
380 }
381
__udpv4_gso_segment_list_csum(struct sk_buff * segs)382 static struct sk_buff *__udpv4_gso_segment_list_csum(struct sk_buff *segs)
383 {
384 struct sk_buff *seg;
385 struct udphdr *uh, *uh2;
386 struct iphdr *iph, *iph2;
387
388 seg = segs;
389 uh = udp_hdr(seg);
390 iph = ip_hdr(seg);
391
392 if ((udp_hdr(seg)->dest == udp_hdr(seg->next)->dest) &&
393 (udp_hdr(seg)->source == udp_hdr(seg->next)->source) &&
394 (ip_hdr(seg)->daddr == ip_hdr(seg->next)->daddr) &&
395 (ip_hdr(seg)->saddr == ip_hdr(seg->next)->saddr))
396 return segs;
397
398 while ((seg = seg->next)) {
399 uh2 = udp_hdr(seg);
400 iph2 = ip_hdr(seg);
401
402 __udpv4_gso_segment_csum(seg,
403 &iph2->saddr, &iph->saddr,
404 &uh2->source, &uh->source);
405 __udpv4_gso_segment_csum(seg,
406 &iph2->daddr, &iph->daddr,
407 &uh2->dest, &uh->dest);
408 }
409
410 return segs;
411 }
412
__udpv6_gso_segment_csum(struct sk_buff * seg,struct in6_addr * oldip,const struct in6_addr * newip,__be16 * oldport,__be16 newport)413 static void __udpv6_gso_segment_csum(struct sk_buff *seg,
414 struct in6_addr *oldip,
415 const struct in6_addr *newip,
416 __be16 *oldport, __be16 newport)
417 {
418 struct udphdr *uh = udp_hdr(seg);
419
420 if (ipv6_addr_equal(oldip, newip) && *oldport == newport)
421 return;
422
423 if (uh->check) {
424 inet_proto_csum_replace16(&uh->check, seg, oldip->s6_addr32,
425 newip->s6_addr32, true);
426
427 inet_proto_csum_replace2(&uh->check, seg, *oldport, newport,
428 false);
429 if (!uh->check)
430 uh->check = CSUM_MANGLED_0;
431 }
432
433 *oldip = *newip;
434 *oldport = newport;
435 }
436
__udpv6_gso_segment_list_csum(struct sk_buff * segs)437 static struct sk_buff *__udpv6_gso_segment_list_csum(struct sk_buff *segs)
438 {
439 const struct ipv6hdr *iph;
440 const struct udphdr *uh;
441 struct ipv6hdr *iph2;
442 struct sk_buff *seg;
443 struct udphdr *uh2;
444
445 seg = segs;
446 uh = udp_hdr(seg);
447 iph = ipv6_hdr(seg);
448 uh2 = udp_hdr(seg->next);
449 iph2 = ipv6_hdr(seg->next);
450
451 if (!(*(const u32 *)&uh->source ^ *(const u32 *)&uh2->source) &&
452 ipv6_addr_equal(&iph->saddr, &iph2->saddr) &&
453 ipv6_addr_equal(&iph->daddr, &iph2->daddr))
454 return segs;
455
456 while ((seg = seg->next)) {
457 uh2 = udp_hdr(seg);
458 iph2 = ipv6_hdr(seg);
459
460 __udpv6_gso_segment_csum(seg, &iph2->saddr, &iph->saddr,
461 &uh2->source, uh->source);
462 __udpv6_gso_segment_csum(seg, &iph2->daddr, &iph->daddr,
463 &uh2->dest, uh->dest);
464 }
465
466 return segs;
467 }
468
__udp_gso_segment_list(struct sk_buff * skb,netdev_features_t features,bool is_ipv6)469 static struct sk_buff *__udp_gso_segment_list(struct sk_buff *skb,
470 netdev_features_t features,
471 bool is_ipv6)
472 {
473 unsigned int mss = skb_shinfo(skb)->gso_size;
474
475 skb = skb_segment_list(skb, features, skb_mac_header_len(skb));
476 if (IS_ERR(skb))
477 return skb;
478
479 udp_hdr(skb)->len = htons(sizeof(struct udphdr) + mss);
480
481 if (is_ipv6)
482 return __udpv6_gso_segment_list_csum(skb);
483 else
484 return __udpv4_gso_segment_list_csum(skb);
485 }
486
__udp_gso_segment(struct sk_buff * gso_skb,netdev_features_t features,bool is_ipv6)487 struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
488 netdev_features_t features, bool is_ipv6)
489 {
490 struct sock *sk = gso_skb->sk;
491 unsigned int sum_truesize = 0;
492 struct sk_buff *segs, *seg;
493 struct udphdr *uh;
494 unsigned int mss;
495 bool copy_dtor;
496 __sum16 check;
497 __be16 newlen;
498 int ret = 0;
499
500 mss = skb_shinfo(gso_skb)->gso_size;
501 if (gso_skb->len <= sizeof(*uh) + mss)
502 return ERR_PTR(-EINVAL);
503
504 if (unlikely(skb_checksum_start(gso_skb) !=
505 skb_transport_header(gso_skb) &&
506 !(skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST)))
507 return ERR_PTR(-EINVAL);
508
509 /* We don't know if egress device can segment and checksum the packet
510 * when IPv6 extension headers are present. Fall back to software GSO.
511 */
512 if (gso_skb->ip_summed != CHECKSUM_PARTIAL)
513 features &= ~(NETIF_F_GSO_UDP_L4 | NETIF_F_CSUM_MASK);
514
515 if (skb_gso_ok(gso_skb, features | NETIF_F_GSO_ROBUST)) {
516 /* Packet is from an untrusted source, reset gso_segs. */
517 skb_shinfo(gso_skb)->gso_segs = DIV_ROUND_UP(gso_skb->len - sizeof(*uh),
518 mss);
519 return NULL;
520 }
521
522 if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST) {
523 /* Detect modified geometry and pass those to skb_segment. */
524 if (skb_pagelen(gso_skb) - sizeof(*uh) == skb_shinfo(gso_skb)->gso_size)
525 return __udp_gso_segment_list(gso_skb, features, is_ipv6);
526
527 ret = __skb_linearize(gso_skb);
528 if (ret)
529 return ERR_PTR(ret);
530
531 /* Setup csum, as fraglist skips this in udp4_gro_receive. */
532 gso_skb->csum_start = skb_transport_header(gso_skb) - gso_skb->head;
533 gso_skb->csum_offset = offsetof(struct udphdr, check);
534 gso_skb->ip_summed = CHECKSUM_PARTIAL;
535
536 uh = udp_hdr(gso_skb);
537 if (is_ipv6)
538 uh->check = ~udp_v6_check(gso_skb->len,
539 &ipv6_hdr(gso_skb)->saddr,
540 &ipv6_hdr(gso_skb)->daddr, 0);
541 else
542 uh->check = ~udp_v4_check(gso_skb->len,
543 ip_hdr(gso_skb)->saddr,
544 ip_hdr(gso_skb)->daddr, 0);
545 }
546
547 skb_pull(gso_skb, sizeof(*uh));
548
549 /* clear destructor to avoid skb_segment assigning it to tail */
550 copy_dtor = gso_skb->destructor == sock_wfree;
551 if (copy_dtor) {
552 gso_skb->destructor = NULL;
553 gso_skb->sk = NULL;
554 }
555
556 segs = skb_segment(gso_skb, features);
557 if (IS_ERR_OR_NULL(segs)) {
558 if (copy_dtor) {
559 gso_skb->destructor = sock_wfree;
560 gso_skb->sk = sk;
561 }
562 return segs;
563 }
564
565 /* GSO partial and frag_list segmentation only requires splitting
566 * the frame into an MSS multiple and possibly a remainder, both
567 * cases return a GSO skb. So update the mss now.
568 */
569 if (skb_is_gso(segs))
570 mss *= skb_shinfo(segs)->gso_segs;
571
572 seg = segs;
573 uh = udp_hdr(seg);
574
575 /* preserve TX timestamp flags and TS key for first segment */
576 skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey;
577 skb_shinfo(seg)->tx_flags |=
578 (skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP);
579
580 /* compute checksum adjustment based on old length versus new */
581 newlen = htons(sizeof(*uh) + mss);
582 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
583
584 for (;;) {
585 if (copy_dtor) {
586 seg->destructor = sock_wfree;
587 seg->sk = sk;
588 sum_truesize += seg->truesize;
589 }
590
591 if (!seg->next)
592 break;
593
594 uh->len = newlen;
595 uh->check = check;
596
597 if (seg->ip_summed == CHECKSUM_PARTIAL)
598 gso_reset_checksum(seg, ~check);
599 else
600 uh->check = gso_make_checksum(seg, ~check) ? :
601 CSUM_MANGLED_0;
602
603 seg = seg->next;
604 uh = udp_hdr(seg);
605 }
606
607 /* last packet can be partial gso_size, account for that in checksum */
608 newlen = htons(skb_tail_pointer(seg) - skb_transport_header(seg) +
609 seg->data_len);
610 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
611
612 uh->len = newlen;
613 uh->check = check;
614
615 if (seg->ip_summed == CHECKSUM_PARTIAL)
616 gso_reset_checksum(seg, ~check);
617 else
618 uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0;
619
620 /* On the TX path, CHECKSUM_NONE and CHECKSUM_UNNECESSARY have the same
621 * meaning. However, check for bad offloads in the GSO stack expects the
622 * latter, if the checksum was calculated in software. To vouch for the
623 * segment skbs we actually need to set it on the gso_skb.
624 */
625 if (gso_skb->ip_summed == CHECKSUM_NONE)
626 gso_skb->ip_summed = CHECKSUM_UNNECESSARY;
627
628 /* update refcount for the packet */
629 if (copy_dtor) {
630 int delta = sum_truesize - gso_skb->truesize;
631
632 /* In some pathological cases, delta can be negative.
633 * We need to either use refcount_add() or refcount_sub_and_test()
634 */
635 if (likely(delta >= 0))
636 refcount_add(delta, &sk->sk_wmem_alloc);
637 else
638 WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));
639 }
640 return segs;
641 }
642 EXPORT_SYMBOL_GPL(__udp_gso_segment);
643
udp4_ufo_fragment(struct sk_buff * skb,netdev_features_t features)644 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
645 netdev_features_t features)
646 {
647 struct sk_buff *segs = ERR_PTR(-EINVAL);
648 unsigned int mss;
649 __wsum csum;
650 struct udphdr *uh;
651 struct iphdr *iph;
652
653 if (skb->encapsulation &&
654 (skb_shinfo(skb)->gso_type &
655 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
656 segs = skb_udp_tunnel_segment(skb, features, false);
657 goto out;
658 }
659
660 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4)))
661 goto out;
662
663 if (!pskb_may_pull(skb, sizeof(struct udphdr)))
664 goto out;
665
666 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
667 return __udp_gso_segment(skb, features, false);
668
669 mss = skb_shinfo(skb)->gso_size;
670 if (unlikely(skb->len <= mss))
671 goto out;
672
673 /* Do software UFO. Complete and fill in the UDP checksum as
674 * HW cannot do checksum of UDP packets sent as multiple
675 * IP fragments.
676 */
677
678 uh = udp_hdr(skb);
679 iph = ip_hdr(skb);
680
681 uh->check = 0;
682 csum = skb_checksum(skb, 0, skb->len, 0);
683 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
684 if (uh->check == 0)
685 uh->check = CSUM_MANGLED_0;
686
687 skb->ip_summed = CHECKSUM_UNNECESSARY;
688
689 /* If there is no outer header we can fake a checksum offload
690 * due to the fact that we have already done the checksum in
691 * software prior to segmenting the frame.
692 */
693 if (!skb->encap_hdr_csum)
694 features |= NETIF_F_HW_CSUM;
695
696 /* Fragment the skb. IP headers of the fragments are updated in
697 * inet_gso_segment()
698 */
699 segs = skb_segment(skb, features);
700 out:
701 return segs;
702 }
703
704
705 #define UDP_GRO_CNT_MAX 64
udp_gro_receive_segment(struct list_head * head,struct sk_buff * skb)706 static struct sk_buff *udp_gro_receive_segment(struct list_head *head,
707 struct sk_buff *skb)
708 {
709 struct udphdr *uh = udp_gro_udphdr(skb);
710 struct sk_buff *pp = NULL;
711 struct udphdr *uh2;
712 struct sk_buff *p;
713 unsigned int ulen;
714 int ret = 0;
715 int flush;
716
717 /* requires non zero csum, for symmetry with GSO */
718 if (!uh->check) {
719 NAPI_GRO_CB(skb)->flush = 1;
720 return NULL;
721 }
722
723 /* Do not deal with padded or malicious packets, sorry ! */
724 ulen = ntohs(uh->len);
725 if (ulen <= sizeof(*uh) || ulen != skb_gro_len(skb)) {
726 NAPI_GRO_CB(skb)->flush = 1;
727 return NULL;
728 }
729 /* pull encapsulating udp header */
730 skb_gro_pull(skb, sizeof(struct udphdr));
731
732 list_for_each_entry(p, head, list) {
733 if (!NAPI_GRO_CB(p)->same_flow)
734 continue;
735
736 uh2 = udp_hdr(p);
737
738 /* Match ports only, as csum is always non zero */
739 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) {
740 NAPI_GRO_CB(p)->same_flow = 0;
741 continue;
742 }
743
744 if (NAPI_GRO_CB(skb)->is_flist != NAPI_GRO_CB(p)->is_flist) {
745 NAPI_GRO_CB(skb)->flush = 1;
746 return p;
747 }
748
749 flush = gro_receive_network_flush(uh, uh2, p);
750
751 /* Terminate the flow on len mismatch or if it grow "too much".
752 * Under small packet flood GRO count could elsewhere grow a lot
753 * leading to excessive truesize values.
754 * On len mismatch merge the first packet shorter than gso_size,
755 * otherwise complete the GRO packet.
756 */
757 if (ulen > ntohs(uh2->len) || flush) {
758 pp = p;
759 } else {
760 if (NAPI_GRO_CB(skb)->is_flist) {
761 if (!pskb_may_pull(skb, skb_gro_offset(skb))) {
762 NAPI_GRO_CB(skb)->flush = 1;
763 return NULL;
764 }
765 if ((skb->ip_summed != p->ip_summed) ||
766 (skb->csum_level != p->csum_level)) {
767 NAPI_GRO_CB(skb)->flush = 1;
768 return NULL;
769 }
770 ret = skb_gro_receive_list(p, skb);
771 } else {
772 skb_gro_postpull_rcsum(skb, uh,
773 sizeof(struct udphdr));
774
775 ret = skb_gro_receive(p, skb);
776 }
777 }
778
779 if (ret || ulen != ntohs(uh2->len) ||
780 NAPI_GRO_CB(p)->count >= UDP_GRO_CNT_MAX)
781 pp = p;
782
783 return pp;
784 }
785
786 /* mismatch, but we never need to flush */
787 return NULL;
788 }
789
udp_gro_receive(struct list_head * head,struct sk_buff * skb,struct udphdr * uh,struct sock * sk)790 struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb,
791 struct udphdr *uh, struct sock *sk)
792 {
793 struct sk_buff *pp = NULL;
794 struct sk_buff *p;
795 struct udphdr *uh2;
796 unsigned int off = skb_gro_offset(skb);
797 int flush = 1;
798
799 /* We can do L4 aggregation only if the packet can't land in a tunnel
800 * otherwise we could corrupt the inner stream. Detecting such packets
801 * cannot be foolproof and the aggregation might still happen in some
802 * cases. Such packets should be caught in udp_unexpected_gso later.
803 */
804 NAPI_GRO_CB(skb)->is_flist = 0;
805 if (!sk || !udp_sk(sk)->gro_receive) {
806 /* If the packet was locally encapsulated in a UDP tunnel that
807 * wasn't detected above, do not GRO.
808 */
809 if (skb->encapsulation)
810 goto out;
811
812 if (skb->dev->features & NETIF_F_GRO_FRAGLIST)
813 NAPI_GRO_CB(skb)->is_flist = sk ? !udp_test_bit(GRO_ENABLED, sk) : 1;
814
815 if ((!sk && (skb->dev->features & NETIF_F_GRO_UDP_FWD)) ||
816 (sk && udp_test_bit(GRO_ENABLED, sk)) || NAPI_GRO_CB(skb)->is_flist)
817 return call_gro_receive(udp_gro_receive_segment, head, skb);
818
819 /* no GRO, be sure flush the current packet */
820 goto out;
821 }
822
823 if (NAPI_GRO_CB(skb)->encap_mark ||
824 (uh->check && skb->ip_summed != CHECKSUM_PARTIAL &&
825 NAPI_GRO_CB(skb)->csum_cnt == 0 &&
826 !NAPI_GRO_CB(skb)->csum_valid))
827 goto out;
828
829 /* mark that this skb passed once through the tunnel gro layer */
830 NAPI_GRO_CB(skb)->encap_mark = 1;
831
832 flush = 0;
833
834 list_for_each_entry(p, head, list) {
835 if (!NAPI_GRO_CB(p)->same_flow)
836 continue;
837
838 uh2 = (struct udphdr *)(p->data + off);
839
840 /* Match ports and either checksums are either both zero
841 * or nonzero.
842 */
843 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
844 (!uh->check ^ !uh2->check)) {
845 NAPI_GRO_CB(p)->same_flow = 0;
846 continue;
847 }
848 }
849
850 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
851 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
852 pp = udp_tunnel_gro_rcv(sk, head, skb);
853
854 out:
855 skb_gro_flush_final(skb, pp, flush);
856 return pp;
857 }
858 EXPORT_SYMBOL(udp_gro_receive);
859
udp4_gro_lookup_skb(struct sk_buff * skb,__be16 sport,__be16 dport)860 static struct sock *udp4_gro_lookup_skb(struct sk_buff *skb, __be16 sport,
861 __be16 dport)
862 {
863 const struct iphdr *iph = skb_gro_network_header(skb);
864 struct net *net = dev_net_rcu(skb->dev);
865 struct sock *sk;
866 int iif, sdif;
867
868 sk = udp_tunnel_sk(net, false);
869 if (sk && dport == htons(sk->sk_num))
870 return sk;
871
872 inet_get_iif_sdif(skb, &iif, &sdif);
873
874 return __udp4_lib_lookup(net, iph->saddr, sport,
875 iph->daddr, dport, iif,
876 sdif, net->ipv4.udp_table, NULL);
877 }
878
879 INDIRECT_CALLABLE_SCOPE
udp4_gro_receive(struct list_head * head,struct sk_buff * skb)880 struct sk_buff *udp4_gro_receive(struct list_head *head, struct sk_buff *skb)
881 {
882 struct udphdr *uh = udp_gro_udphdr(skb);
883 struct sock *sk = NULL;
884 struct sk_buff *pp;
885
886 if (unlikely(!uh))
887 goto flush;
888
889 /* Don't bother verifying checksum if we're going to flush anyway. */
890 if (NAPI_GRO_CB(skb)->flush)
891 goto skip;
892
893 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
894 inet_gro_compute_pseudo))
895 goto flush;
896 else if (uh->check)
897 skb_gro_checksum_try_convert(skb, IPPROTO_UDP,
898 inet_gro_compute_pseudo);
899 skip:
900 NAPI_GRO_CB(skb)->is_ipv6 = 0;
901
902 if (static_branch_unlikely(&udp_encap_needed_key))
903 sk = udp4_gro_lookup_skb(skb, uh->source, uh->dest);
904
905 pp = udp_gro_receive(head, skb, uh, sk);
906 return pp;
907
908 flush:
909 NAPI_GRO_CB(skb)->flush = 1;
910 return NULL;
911 }
912
udp_gro_complete_segment(struct sk_buff * skb)913 static int udp_gro_complete_segment(struct sk_buff *skb)
914 {
915 struct udphdr *uh = udp_hdr(skb);
916
917 skb->csum_start = (unsigned char *)uh - skb->head;
918 skb->csum_offset = offsetof(struct udphdr, check);
919 skb->ip_summed = CHECKSUM_PARTIAL;
920
921 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
922 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_L4;
923
924 if (skb->encapsulation)
925 skb->inner_transport_header = skb->transport_header;
926
927 return 0;
928 }
929
udp_gro_complete(struct sk_buff * skb,int nhoff,udp_lookup_t lookup)930 int udp_gro_complete(struct sk_buff *skb, int nhoff,
931 udp_lookup_t lookup)
932 {
933 __be16 newlen = htons(skb->len - nhoff);
934 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
935 struct sock *sk;
936 int err;
937
938 uh->len = newlen;
939
940 sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb,
941 udp4_lib_lookup_skb, skb, uh->source, uh->dest);
942 if (sk && udp_sk(sk)->gro_complete) {
943 skb_shinfo(skb)->gso_type = uh->check ? SKB_GSO_UDP_TUNNEL_CSUM
944 : SKB_GSO_UDP_TUNNEL;
945
946 /* clear the encap mark, so that inner frag_list gro_complete
947 * can take place
948 */
949 NAPI_GRO_CB(skb)->encap_mark = 0;
950
951 /* Set encapsulation before calling into inner gro_complete()
952 * functions to make them set up the inner offsets.
953 */
954 skb->encapsulation = 1;
955 err = udp_sk(sk)->gro_complete(sk, skb,
956 nhoff + sizeof(struct udphdr));
957 } else {
958 err = udp_gro_complete_segment(skb);
959 }
960
961 if (skb->remcsum_offload)
962 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
963
964 return err;
965 }
966 EXPORT_SYMBOL(udp_gro_complete);
967
udp4_gro_complete(struct sk_buff * skb,int nhoff)968 INDIRECT_CALLABLE_SCOPE int udp4_gro_complete(struct sk_buff *skb, int nhoff)
969 {
970 const u16 offset = NAPI_GRO_CB(skb)->network_offsets[skb->encapsulation];
971 const struct iphdr *iph = (struct iphdr *)(skb->data + offset);
972 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
973
974 /* do fraglist only if there is no outer UDP encap (or we already processed it) */
975 if (NAPI_GRO_CB(skb)->is_flist && !NAPI_GRO_CB(skb)->encap_mark) {
976 uh->len = htons(skb->len - nhoff);
977
978 skb_shinfo(skb)->gso_type |= (SKB_GSO_FRAGLIST|SKB_GSO_UDP_L4);
979 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
980
981 __skb_incr_checksum_unnecessary(skb);
982
983 return 0;
984 }
985
986 if (uh->check)
987 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
988 iph->daddr, 0);
989
990 return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
991 }
992
udpv4_offload_init(void)993 int __init udpv4_offload_init(void)
994 {
995 net_hotdata.udpv4_offload = (struct net_offload) {
996 .callbacks = {
997 .gso_segment = udp4_ufo_fragment,
998 .gro_receive = udp4_gro_receive,
999 .gro_complete = udp4_gro_complete,
1000 },
1001 };
1002
1003 udp_tunnel_gro_init();
1004 return inet_add_offload(&net_hotdata.udpv4_offload, IPPROTO_UDP);
1005 }
1006