Loading...
1/*
2 * IPV4 GSO/GRO offload support
3 * Linux INET implementation
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License
7 * as published by the Free Software Foundation; either version
8 * 2 of the License, or (at your option) any later version.
9 *
10 * UDPv4 GSO support
11 */
12
13#include <linux/skbuff.h>
14#include <net/udp.h>
15#include <net/protocol.h>
16
17static struct sk_buff *__skb_udp_tunnel_segment(struct sk_buff *skb,
18 netdev_features_t features,
19 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
20 netdev_features_t features),
21 __be16 new_protocol, bool is_ipv6)
22{
23 int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
24 bool remcsum, need_csum, offload_csum, gso_partial;
25 struct sk_buff *segs = ERR_PTR(-EINVAL);
26 struct udphdr *uh = udp_hdr(skb);
27 u16 mac_offset = skb->mac_header;
28 __be16 protocol = skb->protocol;
29 u16 mac_len = skb->mac_len;
30 int udp_offset, outer_hlen;
31 __wsum partial;
32 bool need_ipsec;
33
34 if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
35 goto out;
36
37 /* Adjust partial header checksum to negate old length.
38 * We cannot rely on the value contained in uh->len as it is
39 * possible that the actual value exceeds the boundaries of the
40 * 16 bit length field due to the header being added outside of an
41 * IP or IPv6 frame that was already limited to 64K - 1.
42 */
43 if (skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)
44 partial = (__force __wsum)uh->len;
45 else
46 partial = (__force __wsum)htonl(skb->len);
47 partial = csum_sub(csum_unfold(uh->check), partial);
48
49 /* setup inner skb. */
50 skb->encapsulation = 0;
51 SKB_GSO_CB(skb)->encap_level = 0;
52 __skb_pull(skb, tnl_hlen);
53 skb_reset_mac_header(skb);
54 skb_set_network_header(skb, skb_inner_network_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));
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
74 /* The only checksum offload we care about from here on out is the
75 * outer one so strip the existing checksum feature flags and
76 * instead set the flag based on our outer checksum offload value.
77 */
78 if (remcsum) {
79 features &= ~NETIF_F_CSUM_MASK;
80 if (!need_csum || offload_csum)
81 features |= NETIF_F_HW_CSUM;
82 }
83
84 /* segment inner packet. */
85 segs = gso_inner_segment(skb, features);
86 if (IS_ERR_OR_NULL(segs)) {
87 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
88 mac_len);
89 goto out;
90 }
91
92 gso_partial = !!(skb_shinfo(segs)->gso_type & SKB_GSO_PARTIAL);
93
94 outer_hlen = skb_tnl_header_len(skb);
95 udp_offset = outer_hlen - tnl_hlen;
96 skb = segs;
97 do {
98 unsigned int len;
99
100 if (remcsum)
101 skb->ip_summed = CHECKSUM_NONE;
102
103 /* Set up inner headers if we are offloading inner checksum */
104 if (skb->ip_summed == CHECKSUM_PARTIAL) {
105 skb_reset_inner_headers(skb);
106 skb->encapsulation = 1;
107 }
108
109 skb->mac_len = mac_len;
110 skb->protocol = protocol;
111
112 __skb_push(skb, outer_hlen);
113 skb_reset_mac_header(skb);
114 skb_set_network_header(skb, mac_len);
115 skb_set_transport_header(skb, udp_offset);
116 len = skb->len - udp_offset;
117 uh = udp_hdr(skb);
118
119 /* If we are only performing partial GSO the inner header
120 * will be using a length value equal to only one MSS sized
121 * segment instead of the entire frame.
122 */
123 if (gso_partial && skb_is_gso(skb)) {
124 uh->len = htons(skb_shinfo(skb)->gso_size +
125 SKB_GSO_CB(skb)->data_offset +
126 skb->head - (unsigned char *)uh);
127 } else {
128 uh->len = htons(len);
129 }
130
131 if (!need_csum)
132 continue;
133
134 uh->check = ~csum_fold(csum_add(partial,
135 (__force __wsum)htonl(len)));
136
137 if (skb->encapsulation || !offload_csum) {
138 uh->check = gso_make_checksum(skb, ~uh->check);
139 if (uh->check == 0)
140 uh->check = CSUM_MANGLED_0;
141 } else {
142 skb->ip_summed = CHECKSUM_PARTIAL;
143 skb->csum_start = skb_transport_header(skb) - skb->head;
144 skb->csum_offset = offsetof(struct udphdr, check);
145 }
146 } while ((skb = skb->next));
147out:
148 return segs;
149}
150
151struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
152 netdev_features_t features,
153 bool is_ipv6)
154{
155 __be16 protocol = skb->protocol;
156 const struct net_offload **offloads;
157 const struct net_offload *ops;
158 struct sk_buff *segs = ERR_PTR(-EINVAL);
159 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
160 netdev_features_t features);
161
162 rcu_read_lock();
163
164 switch (skb->inner_protocol_type) {
165 case ENCAP_TYPE_ETHER:
166 protocol = skb->inner_protocol;
167 gso_inner_segment = skb_mac_gso_segment;
168 break;
169 case ENCAP_TYPE_IPPROTO:
170 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
171 ops = rcu_dereference(offloads[skb->inner_ipproto]);
172 if (!ops || !ops->callbacks.gso_segment)
173 goto out_unlock;
174 gso_inner_segment = ops->callbacks.gso_segment;
175 break;
176 default:
177 goto out_unlock;
178 }
179
180 segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
181 protocol, is_ipv6);
182
183out_unlock:
184 rcu_read_unlock();
185
186 return segs;
187}
188EXPORT_SYMBOL(skb_udp_tunnel_segment);
189
190static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
191 netdev_features_t features)
192{
193 struct sk_buff *segs = ERR_PTR(-EINVAL);
194 unsigned int mss;
195 __wsum csum;
196 struct udphdr *uh;
197 struct iphdr *iph;
198
199 if (skb->encapsulation &&
200 (skb_shinfo(skb)->gso_type &
201 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
202 segs = skb_udp_tunnel_segment(skb, features, false);
203 goto out;
204 }
205
206 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP))
207 goto out;
208
209 if (!pskb_may_pull(skb, sizeof(struct udphdr)))
210 goto out;
211
212 mss = skb_shinfo(skb)->gso_size;
213 if (unlikely(skb->len <= mss))
214 goto out;
215
216 /* Do software UFO. Complete and fill in the UDP checksum as
217 * HW cannot do checksum of UDP packets sent as multiple
218 * IP fragments.
219 */
220
221 uh = udp_hdr(skb);
222 iph = ip_hdr(skb);
223
224 uh->check = 0;
225 csum = skb_checksum(skb, 0, skb->len, 0);
226 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
227 if (uh->check == 0)
228 uh->check = CSUM_MANGLED_0;
229
230 skb->ip_summed = CHECKSUM_UNNECESSARY;
231
232 /* If there is no outer header we can fake a checksum offload
233 * due to the fact that we have already done the checksum in
234 * software prior to segmenting the frame.
235 */
236 if (!skb->encap_hdr_csum)
237 features |= NETIF_F_HW_CSUM;
238
239 /* Fragment the skb. IP headers of the fragments are updated in
240 * inet_gso_segment()
241 */
242 segs = skb_segment(skb, features);
243out:
244 return segs;
245}
246
247struct sk_buff **udp_gro_receive(struct sk_buff **head, struct sk_buff *skb,
248 struct udphdr *uh, udp_lookup_t lookup)
249{
250 struct sk_buff *p, **pp = NULL;
251 struct udphdr *uh2;
252 unsigned int off = skb_gro_offset(skb);
253 int flush = 1;
254 struct sock *sk;
255
256 if (NAPI_GRO_CB(skb)->encap_mark ||
257 (skb->ip_summed != CHECKSUM_PARTIAL &&
258 NAPI_GRO_CB(skb)->csum_cnt == 0 &&
259 !NAPI_GRO_CB(skb)->csum_valid))
260 goto out;
261
262 /* mark that this skb passed once through the tunnel gro layer */
263 NAPI_GRO_CB(skb)->encap_mark = 1;
264
265 rcu_read_lock();
266 sk = (*lookup)(skb, uh->source, uh->dest);
267
268 if (sk && udp_sk(sk)->gro_receive)
269 goto unflush;
270 goto out_unlock;
271
272unflush:
273 flush = 0;
274
275 for (p = *head; p; p = p->next) {
276 if (!NAPI_GRO_CB(p)->same_flow)
277 continue;
278
279 uh2 = (struct udphdr *)(p->data + off);
280
281 /* Match ports and either checksums are either both zero
282 * or nonzero.
283 */
284 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
285 (!uh->check ^ !uh2->check)) {
286 NAPI_GRO_CB(p)->same_flow = 0;
287 continue;
288 }
289 }
290
291 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
292 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
293 pp = call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
294
295out_unlock:
296 rcu_read_unlock();
297out:
298 NAPI_GRO_CB(skb)->flush |= flush;
299 return pp;
300}
301EXPORT_SYMBOL(udp_gro_receive);
302
303static struct sk_buff **udp4_gro_receive(struct sk_buff **head,
304 struct sk_buff *skb)
305{
306 struct udphdr *uh = udp_gro_udphdr(skb);
307
308 if (unlikely(!uh))
309 goto flush;
310
311 /* Don't bother verifying checksum if we're going to flush anyway. */
312 if (NAPI_GRO_CB(skb)->flush)
313 goto skip;
314
315 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
316 inet_gro_compute_pseudo))
317 goto flush;
318 else if (uh->check)
319 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, uh->check,
320 inet_gro_compute_pseudo);
321skip:
322 NAPI_GRO_CB(skb)->is_ipv6 = 0;
323 return udp_gro_receive(head, skb, uh, udp4_lib_lookup_skb);
324
325flush:
326 NAPI_GRO_CB(skb)->flush = 1;
327 return NULL;
328}
329
330int udp_gro_complete(struct sk_buff *skb, int nhoff,
331 udp_lookup_t lookup)
332{
333 __be16 newlen = htons(skb->len - nhoff);
334 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
335 int err = -ENOSYS;
336 struct sock *sk;
337
338 uh->len = newlen;
339
340 /* Set encapsulation before calling into inner gro_complete() functions
341 * to make them set up the inner offsets.
342 */
343 skb->encapsulation = 1;
344
345 rcu_read_lock();
346 sk = (*lookup)(skb, uh->source, uh->dest);
347 if (sk && udp_sk(sk)->gro_complete)
348 err = udp_sk(sk)->gro_complete(sk, skb,
349 nhoff + sizeof(struct udphdr));
350 rcu_read_unlock();
351
352 if (skb->remcsum_offload)
353 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
354
355 return err;
356}
357EXPORT_SYMBOL(udp_gro_complete);
358
359static int udp4_gro_complete(struct sk_buff *skb, int nhoff)
360{
361 const struct iphdr *iph = ip_hdr(skb);
362 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
363
364 if (uh->check) {
365 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL_CSUM;
366 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
367 iph->daddr, 0);
368 } else {
369 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL;
370 }
371
372 return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
373}
374
375static const struct net_offload udpv4_offload = {
376 .callbacks = {
377 .gso_segment = udp4_ufo_fragment,
378 .gro_receive = udp4_gro_receive,
379 .gro_complete = udp4_gro_complete,
380 },
381};
382
383int __init udpv4_offload_init(void)
384{
385 return inet_add_offload(&udpv4_offload, IPPROTO_UDP);
386}
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
16static 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));
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));
149out:
150 return segs;
151}
152
153struct 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
185out_unlock:
186 rcu_read_unlock();
187
188 return segs;
189}
190EXPORT_SYMBOL(skb_udp_tunnel_segment);
191
192static 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
219static 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
250static struct sk_buff *__udp_gso_segment_list(struct sk_buff *skb,
251 netdev_features_t features,
252 bool is_ipv6)
253{
254 unsigned int mss = skb_shinfo(skb)->gso_size;
255
256 skb = skb_segment_list(skb, features, skb_mac_header_len(skb));
257 if (IS_ERR(skb))
258 return skb;
259
260 udp_hdr(skb)->len = htons(sizeof(struct udphdr) + mss);
261
262 return is_ipv6 ? skb : __udpv4_gso_segment_list_csum(skb);
263}
264
265struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
266 netdev_features_t features, bool is_ipv6)
267{
268 struct sock *sk = gso_skb->sk;
269 unsigned int sum_truesize = 0;
270 struct sk_buff *segs, *seg;
271 struct udphdr *uh;
272 unsigned int mss;
273 bool copy_dtor;
274 __sum16 check;
275 __be16 newlen;
276
277 mss = skb_shinfo(gso_skb)->gso_size;
278 if (gso_skb->len <= sizeof(*uh) + mss)
279 return ERR_PTR(-EINVAL);
280
281 if (unlikely(skb_checksum_start(gso_skb) !=
282 skb_transport_header(gso_skb) &&
283 !(skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST)))
284 return ERR_PTR(-EINVAL);
285
286 /* We don't know if egress device can segment and checksum the packet
287 * when IPv6 extension headers are present. Fall back to software GSO.
288 */
289 if (gso_skb->ip_summed != CHECKSUM_PARTIAL)
290 features &= ~(NETIF_F_GSO_UDP_L4 | NETIF_F_CSUM_MASK);
291
292 if (skb_gso_ok(gso_skb, features | NETIF_F_GSO_ROBUST)) {
293 /* Packet is from an untrusted source, reset gso_segs. */
294 skb_shinfo(gso_skb)->gso_segs = DIV_ROUND_UP(gso_skb->len - sizeof(*uh),
295 mss);
296 return NULL;
297 }
298
299 if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST) {
300 /* Detect modified geometry and pass those to skb_segment. */
301 if (skb_pagelen(gso_skb) - sizeof(*uh) == skb_shinfo(gso_skb)->gso_size)
302 return __udp_gso_segment_list(gso_skb, features, is_ipv6);
303
304 /* Setup csum, as fraglist skips this in udp4_gro_receive. */
305 gso_skb->csum_start = skb_transport_header(gso_skb) - gso_skb->head;
306 gso_skb->csum_offset = offsetof(struct udphdr, check);
307 gso_skb->ip_summed = CHECKSUM_PARTIAL;
308
309 uh = udp_hdr(gso_skb);
310 if (is_ipv6)
311 uh->check = ~udp_v6_check(gso_skb->len,
312 &ipv6_hdr(gso_skb)->saddr,
313 &ipv6_hdr(gso_skb)->daddr, 0);
314 else
315 uh->check = ~udp_v4_check(gso_skb->len,
316 ip_hdr(gso_skb)->saddr,
317 ip_hdr(gso_skb)->daddr, 0);
318 }
319
320 skb_pull(gso_skb, sizeof(*uh));
321
322 /* clear destructor to avoid skb_segment assigning it to tail */
323 copy_dtor = gso_skb->destructor == sock_wfree;
324 if (copy_dtor) {
325 gso_skb->destructor = NULL;
326 gso_skb->sk = NULL;
327 }
328
329 segs = skb_segment(gso_skb, features);
330 if (IS_ERR_OR_NULL(segs)) {
331 if (copy_dtor) {
332 gso_skb->destructor = sock_wfree;
333 gso_skb->sk = sk;
334 }
335 return segs;
336 }
337
338 /* GSO partial and frag_list segmentation only requires splitting
339 * the frame into an MSS multiple and possibly a remainder, both
340 * cases return a GSO skb. So update the mss now.
341 */
342 if (skb_is_gso(segs))
343 mss *= skb_shinfo(segs)->gso_segs;
344
345 seg = segs;
346 uh = udp_hdr(seg);
347
348 /* preserve TX timestamp flags and TS key for first segment */
349 skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey;
350 skb_shinfo(seg)->tx_flags |=
351 (skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP);
352
353 /* compute checksum adjustment based on old length versus new */
354 newlen = htons(sizeof(*uh) + mss);
355 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
356
357 for (;;) {
358 if (copy_dtor) {
359 seg->destructor = sock_wfree;
360 seg->sk = sk;
361 sum_truesize += seg->truesize;
362 }
363
364 if (!seg->next)
365 break;
366
367 uh->len = newlen;
368 uh->check = check;
369
370 if (seg->ip_summed == CHECKSUM_PARTIAL)
371 gso_reset_checksum(seg, ~check);
372 else
373 uh->check = gso_make_checksum(seg, ~check) ? :
374 CSUM_MANGLED_0;
375
376 seg = seg->next;
377 uh = udp_hdr(seg);
378 }
379
380 /* last packet can be partial gso_size, account for that in checksum */
381 newlen = htons(skb_tail_pointer(seg) - skb_transport_header(seg) +
382 seg->data_len);
383 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
384
385 uh->len = newlen;
386 uh->check = check;
387
388 if (seg->ip_summed == CHECKSUM_PARTIAL)
389 gso_reset_checksum(seg, ~check);
390 else
391 uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0;
392
393 /* On the TX path, CHECKSUM_NONE and CHECKSUM_UNNECESSARY have the same
394 * meaning. However, check for bad offloads in the GSO stack expects the
395 * latter, if the checksum was calculated in software. To vouch for the
396 * segment skbs we actually need to set it on the gso_skb.
397 */
398 if (gso_skb->ip_summed == CHECKSUM_NONE)
399 gso_skb->ip_summed = CHECKSUM_UNNECESSARY;
400
401 /* update refcount for the packet */
402 if (copy_dtor) {
403 int delta = sum_truesize - gso_skb->truesize;
404
405 /* In some pathological cases, delta can be negative.
406 * We need to either use refcount_add() or refcount_sub_and_test()
407 */
408 if (likely(delta >= 0))
409 refcount_add(delta, &sk->sk_wmem_alloc);
410 else
411 WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));
412 }
413 return segs;
414}
415EXPORT_SYMBOL_GPL(__udp_gso_segment);
416
417static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
418 netdev_features_t features)
419{
420 struct sk_buff *segs = ERR_PTR(-EINVAL);
421 unsigned int mss;
422 __wsum csum;
423 struct udphdr *uh;
424 struct iphdr *iph;
425
426 if (skb->encapsulation &&
427 (skb_shinfo(skb)->gso_type &
428 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
429 segs = skb_udp_tunnel_segment(skb, features, false);
430 goto out;
431 }
432
433 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4)))
434 goto out;
435
436 if (!pskb_may_pull(skb, sizeof(struct udphdr)))
437 goto out;
438
439 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
440 return __udp_gso_segment(skb, features, false);
441
442 mss = skb_shinfo(skb)->gso_size;
443 if (unlikely(skb->len <= mss))
444 goto out;
445
446 /* Do software UFO. Complete and fill in the UDP checksum as
447 * HW cannot do checksum of UDP packets sent as multiple
448 * IP fragments.
449 */
450
451 uh = udp_hdr(skb);
452 iph = ip_hdr(skb);
453
454 uh->check = 0;
455 csum = skb_checksum(skb, 0, skb->len, 0);
456 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
457 if (uh->check == 0)
458 uh->check = CSUM_MANGLED_0;
459
460 skb->ip_summed = CHECKSUM_UNNECESSARY;
461
462 /* If there is no outer header we can fake a checksum offload
463 * due to the fact that we have already done the checksum in
464 * software prior to segmenting the frame.
465 */
466 if (!skb->encap_hdr_csum)
467 features |= NETIF_F_HW_CSUM;
468
469 /* Fragment the skb. IP headers of the fragments are updated in
470 * inet_gso_segment()
471 */
472 segs = skb_segment(skb, features);
473out:
474 return segs;
475}
476
477
478#define UDP_GRO_CNT_MAX 64
479static struct sk_buff *udp_gro_receive_segment(struct list_head *head,
480 struct sk_buff *skb)
481{
482 struct udphdr *uh = udp_gro_udphdr(skb);
483 struct sk_buff *pp = NULL;
484 struct udphdr *uh2;
485 struct sk_buff *p;
486 unsigned int ulen;
487 int ret = 0;
488 int flush;
489
490 /* requires non zero csum, for symmetry with GSO */
491 if (!uh->check) {
492 NAPI_GRO_CB(skb)->flush = 1;
493 return NULL;
494 }
495
496 /* Do not deal with padded or malicious packets, sorry ! */
497 ulen = ntohs(uh->len);
498 if (ulen <= sizeof(*uh) || ulen != skb_gro_len(skb)) {
499 NAPI_GRO_CB(skb)->flush = 1;
500 return NULL;
501 }
502 /* pull encapsulating udp header */
503 skb_gro_pull(skb, sizeof(struct udphdr));
504
505 list_for_each_entry(p, head, list) {
506 if (!NAPI_GRO_CB(p)->same_flow)
507 continue;
508
509 uh2 = udp_hdr(p);
510
511 /* Match ports only, as csum is always non zero */
512 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) {
513 NAPI_GRO_CB(p)->same_flow = 0;
514 continue;
515 }
516
517 if (NAPI_GRO_CB(skb)->is_flist != NAPI_GRO_CB(p)->is_flist) {
518 NAPI_GRO_CB(skb)->flush = 1;
519 return p;
520 }
521
522 flush = gro_receive_network_flush(uh, uh2, p);
523
524 /* Terminate the flow on len mismatch or if it grow "too much".
525 * Under small packet flood GRO count could elsewhere grow a lot
526 * leading to excessive truesize values.
527 * On len mismatch merge the first packet shorter than gso_size,
528 * otherwise complete the GRO packet.
529 */
530 if (ulen > ntohs(uh2->len) || flush) {
531 pp = p;
532 } else {
533 if (NAPI_GRO_CB(skb)->is_flist) {
534 if (!pskb_may_pull(skb, skb_gro_offset(skb))) {
535 NAPI_GRO_CB(skb)->flush = 1;
536 return NULL;
537 }
538 if ((skb->ip_summed != p->ip_summed) ||
539 (skb->csum_level != p->csum_level)) {
540 NAPI_GRO_CB(skb)->flush = 1;
541 return NULL;
542 }
543 ret = skb_gro_receive_list(p, skb);
544 } else {
545 skb_gro_postpull_rcsum(skb, uh,
546 sizeof(struct udphdr));
547
548 ret = skb_gro_receive(p, skb);
549 }
550 }
551
552 if (ret || ulen != ntohs(uh2->len) ||
553 NAPI_GRO_CB(p)->count >= UDP_GRO_CNT_MAX)
554 pp = p;
555
556 return pp;
557 }
558
559 /* mismatch, but we never need to flush */
560 return NULL;
561}
562
563struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb,
564 struct udphdr *uh, struct sock *sk)
565{
566 struct sk_buff *pp = NULL;
567 struct sk_buff *p;
568 struct udphdr *uh2;
569 unsigned int off = skb_gro_offset(skb);
570 int flush = 1;
571
572 /* We can do L4 aggregation only if the packet can't land in a tunnel
573 * otherwise we could corrupt the inner stream. Detecting such packets
574 * cannot be foolproof and the aggregation might still happen in some
575 * cases. Such packets should be caught in udp_unexpected_gso later.
576 */
577 NAPI_GRO_CB(skb)->is_flist = 0;
578 if (!sk || !udp_sk(sk)->gro_receive) {
579 /* If the packet was locally encapsulated in a UDP tunnel that
580 * wasn't detected above, do not GRO.
581 */
582 if (skb->encapsulation)
583 goto out;
584
585 if (skb->dev->features & NETIF_F_GRO_FRAGLIST)
586 NAPI_GRO_CB(skb)->is_flist = sk ? !udp_test_bit(GRO_ENABLED, sk) : 1;
587
588 if ((!sk && (skb->dev->features & NETIF_F_GRO_UDP_FWD)) ||
589 (sk && udp_test_bit(GRO_ENABLED, sk)) || NAPI_GRO_CB(skb)->is_flist)
590 return call_gro_receive(udp_gro_receive_segment, head, skb);
591
592 /* no GRO, be sure flush the current packet */
593 goto out;
594 }
595
596 if (NAPI_GRO_CB(skb)->encap_mark ||
597 (uh->check && skb->ip_summed != CHECKSUM_PARTIAL &&
598 NAPI_GRO_CB(skb)->csum_cnt == 0 &&
599 !NAPI_GRO_CB(skb)->csum_valid))
600 goto out;
601
602 /* mark that this skb passed once through the tunnel gro layer */
603 NAPI_GRO_CB(skb)->encap_mark = 1;
604
605 flush = 0;
606
607 list_for_each_entry(p, head, list) {
608 if (!NAPI_GRO_CB(p)->same_flow)
609 continue;
610
611 uh2 = (struct udphdr *)(p->data + off);
612
613 /* Match ports and either checksums are either both zero
614 * or nonzero.
615 */
616 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
617 (!uh->check ^ !uh2->check)) {
618 NAPI_GRO_CB(p)->same_flow = 0;
619 continue;
620 }
621 }
622
623 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
624 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
625 pp = call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
626
627out:
628 skb_gro_flush_final(skb, pp, flush);
629 return pp;
630}
631EXPORT_SYMBOL(udp_gro_receive);
632
633static struct sock *udp4_gro_lookup_skb(struct sk_buff *skb, __be16 sport,
634 __be16 dport)
635{
636 const struct iphdr *iph = skb_gro_network_header(skb);
637 struct net *net = dev_net(skb->dev);
638 int iif, sdif;
639
640 inet_get_iif_sdif(skb, &iif, &sdif);
641
642 return __udp4_lib_lookup(net, iph->saddr, sport,
643 iph->daddr, dport, iif,
644 sdif, net->ipv4.udp_table, NULL);
645}
646
647INDIRECT_CALLABLE_SCOPE
648struct sk_buff *udp4_gro_receive(struct list_head *head, struct sk_buff *skb)
649{
650 struct udphdr *uh = udp_gro_udphdr(skb);
651 struct sock *sk = NULL;
652 struct sk_buff *pp;
653
654 if (unlikely(!uh))
655 goto flush;
656
657 /* Don't bother verifying checksum if we're going to flush anyway. */
658 if (NAPI_GRO_CB(skb)->flush)
659 goto skip;
660
661 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
662 inet_gro_compute_pseudo))
663 goto flush;
664 else if (uh->check)
665 skb_gro_checksum_try_convert(skb, IPPROTO_UDP,
666 inet_gro_compute_pseudo);
667skip:
668 NAPI_GRO_CB(skb)->is_ipv6 = 0;
669
670 if (static_branch_unlikely(&udp_encap_needed_key))
671 sk = udp4_gro_lookup_skb(skb, uh->source, uh->dest);
672
673 pp = udp_gro_receive(head, skb, uh, sk);
674 return pp;
675
676flush:
677 NAPI_GRO_CB(skb)->flush = 1;
678 return NULL;
679}
680
681static int udp_gro_complete_segment(struct sk_buff *skb)
682{
683 struct udphdr *uh = udp_hdr(skb);
684
685 skb->csum_start = (unsigned char *)uh - skb->head;
686 skb->csum_offset = offsetof(struct udphdr, check);
687 skb->ip_summed = CHECKSUM_PARTIAL;
688
689 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
690 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_L4;
691
692 if (skb->encapsulation)
693 skb->inner_transport_header = skb->transport_header;
694
695 return 0;
696}
697
698int udp_gro_complete(struct sk_buff *skb, int nhoff,
699 udp_lookup_t lookup)
700{
701 __be16 newlen = htons(skb->len - nhoff);
702 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
703 struct sock *sk;
704 int err;
705
706 uh->len = newlen;
707
708 sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb,
709 udp4_lib_lookup_skb, skb, uh->source, uh->dest);
710 if (sk && udp_sk(sk)->gro_complete) {
711 skb_shinfo(skb)->gso_type = uh->check ? SKB_GSO_UDP_TUNNEL_CSUM
712 : SKB_GSO_UDP_TUNNEL;
713
714 /* clear the encap mark, so that inner frag_list gro_complete
715 * can take place
716 */
717 NAPI_GRO_CB(skb)->encap_mark = 0;
718
719 /* Set encapsulation before calling into inner gro_complete()
720 * functions to make them set up the inner offsets.
721 */
722 skb->encapsulation = 1;
723 err = udp_sk(sk)->gro_complete(sk, skb,
724 nhoff + sizeof(struct udphdr));
725 } else {
726 err = udp_gro_complete_segment(skb);
727 }
728
729 if (skb->remcsum_offload)
730 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
731
732 return err;
733}
734EXPORT_SYMBOL(udp_gro_complete);
735
736INDIRECT_CALLABLE_SCOPE int udp4_gro_complete(struct sk_buff *skb, int nhoff)
737{
738 const u16 offset = NAPI_GRO_CB(skb)->network_offsets[skb->encapsulation];
739 const struct iphdr *iph = (struct iphdr *)(skb->data + offset);
740 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
741
742 /* do fraglist only if there is no outer UDP encap (or we already processed it) */
743 if (NAPI_GRO_CB(skb)->is_flist && !NAPI_GRO_CB(skb)->encap_mark) {
744 uh->len = htons(skb->len - nhoff);
745
746 skb_shinfo(skb)->gso_type |= (SKB_GSO_FRAGLIST|SKB_GSO_UDP_L4);
747 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
748
749 __skb_incr_checksum_unnecessary(skb);
750
751 return 0;
752 }
753
754 if (uh->check)
755 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
756 iph->daddr, 0);
757
758 return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
759}
760
761int __init udpv4_offload_init(void)
762{
763 net_hotdata.udpv4_offload = (struct net_offload) {
764 .callbacks = {
765 .gso_segment = udp4_ufo_fragment,
766 .gro_receive = udp4_gro_receive,
767 .gro_complete = udp4_gro_complete,
768 },
769 };
770 return inet_add_offload(&net_hotdata.udpv4_offload, IPPROTO_UDP);
771}