xref: /openbmc/linux/net/ipv4/udp_offload.c (revision 46aa2f30)
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 
17 static 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, ufo;
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 
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 	partial = csum_sub(csum_unfold(uh->check),
43 			   (__force __wsum)htonl(skb->len));
44 
45 	/* setup inner skb. */
46 	skb->encapsulation = 0;
47 	SKB_GSO_CB(skb)->encap_level = 0;
48 	__skb_pull(skb, tnl_hlen);
49 	skb_reset_mac_header(skb);
50 	skb_set_network_header(skb, skb_inner_network_offset(skb));
51 	skb->mac_len = skb_inner_network_offset(skb);
52 	skb->protocol = new_protocol;
53 
54 	need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
55 	skb->encap_hdr_csum = need_csum;
56 
57 	remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
58 	skb->remcsum_offload = remcsum;
59 
60 	ufo = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP);
61 
62 	/* Try to offload checksum if possible */
63 	offload_csum = !!(need_csum &&
64 			  (skb->dev->features &
65 			   (is_ipv6 ? (NETIF_F_HW_CSUM | NETIF_F_IPV6_CSUM) :
66 				      (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM))));
67 
68 	features &= skb->dev->hw_enc_features;
69 
70 	/* The only checksum offload we care about from here on out is the
71 	 * outer one so strip the existing checksum feature flags and
72 	 * instead set the flag based on our outer checksum offload value.
73 	 */
74 	if (remcsum || ufo) {
75 		features &= ~NETIF_F_CSUM_MASK;
76 		if (!need_csum || offload_csum)
77 			features |= NETIF_F_HW_CSUM;
78 	}
79 
80 	/* segment inner packet. */
81 	segs = gso_inner_segment(skb, features);
82 	if (IS_ERR_OR_NULL(segs)) {
83 		skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
84 				     mac_len);
85 		goto out;
86 	}
87 
88 	outer_hlen = skb_tnl_header_len(skb);
89 	udp_offset = outer_hlen - tnl_hlen;
90 	skb = segs;
91 	do {
92 		__be16 len;
93 
94 		if (remcsum)
95 			skb->ip_summed = CHECKSUM_NONE;
96 
97 		/* Set up inner headers if we are offloading inner checksum */
98 		if (skb->ip_summed == CHECKSUM_PARTIAL) {
99 			skb_reset_inner_headers(skb);
100 			skb->encapsulation = 1;
101 		}
102 
103 		skb->mac_len = mac_len;
104 		skb->protocol = protocol;
105 
106 		__skb_push(skb, outer_hlen);
107 		skb_reset_mac_header(skb);
108 		skb_set_network_header(skb, mac_len);
109 		skb_set_transport_header(skb, udp_offset);
110 		len = htons(skb->len - udp_offset);
111 		uh = udp_hdr(skb);
112 		uh->len = len;
113 
114 		if (!need_csum)
115 			continue;
116 
117 		uh->check = ~csum_fold(csum_add(partial, (__force __wsum)len));
118 
119 		if (skb->encapsulation || !offload_csum) {
120 			uh->check = gso_make_checksum(skb, ~uh->check);
121 			if (uh->check == 0)
122 				uh->check = CSUM_MANGLED_0;
123 		} else {
124 			skb->ip_summed = CHECKSUM_PARTIAL;
125 			skb->csum_start = skb_transport_header(skb) - skb->head;
126 			skb->csum_offset = offsetof(struct udphdr, check);
127 		}
128 	} while ((skb = skb->next));
129 out:
130 	return segs;
131 }
132 
133 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
134 				       netdev_features_t features,
135 				       bool is_ipv6)
136 {
137 	__be16 protocol = skb->protocol;
138 	const struct net_offload **offloads;
139 	const struct net_offload *ops;
140 	struct sk_buff *segs = ERR_PTR(-EINVAL);
141 	struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
142 					     netdev_features_t features);
143 
144 	rcu_read_lock();
145 
146 	switch (skb->inner_protocol_type) {
147 	case ENCAP_TYPE_ETHER:
148 		protocol = skb->inner_protocol;
149 		gso_inner_segment = skb_mac_gso_segment;
150 		break;
151 	case ENCAP_TYPE_IPPROTO:
152 		offloads = is_ipv6 ? inet6_offloads : inet_offloads;
153 		ops = rcu_dereference(offloads[skb->inner_ipproto]);
154 		if (!ops || !ops->callbacks.gso_segment)
155 			goto out_unlock;
156 		gso_inner_segment = ops->callbacks.gso_segment;
157 		break;
158 	default:
159 		goto out_unlock;
160 	}
161 
162 	segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
163 					protocol, is_ipv6);
164 
165 out_unlock:
166 	rcu_read_unlock();
167 
168 	return segs;
169 }
170 EXPORT_SYMBOL(skb_udp_tunnel_segment);
171 
172 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
173 					 netdev_features_t features)
174 {
175 	struct sk_buff *segs = ERR_PTR(-EINVAL);
176 	unsigned int mss;
177 	__wsum csum;
178 	struct udphdr *uh;
179 	struct iphdr *iph;
180 
181 	if (skb->encapsulation &&
182 	    (skb_shinfo(skb)->gso_type &
183 	     (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
184 		segs = skb_udp_tunnel_segment(skb, features, false);
185 		goto out;
186 	}
187 
188 	if (!pskb_may_pull(skb, sizeof(struct udphdr)))
189 		goto out;
190 
191 	mss = skb_shinfo(skb)->gso_size;
192 	if (unlikely(skb->len <= mss))
193 		goto out;
194 
195 	if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
196 		/* Packet is from an untrusted source, reset gso_segs. */
197 		int type = skb_shinfo(skb)->gso_type;
198 
199 		if (unlikely(type & ~(SKB_GSO_UDP | SKB_GSO_DODGY |
200 				      SKB_GSO_UDP_TUNNEL |
201 				      SKB_GSO_UDP_TUNNEL_CSUM |
202 				      SKB_GSO_TUNNEL_REMCSUM |
203 				      SKB_GSO_IPIP |
204 				      SKB_GSO_GRE | SKB_GSO_GRE_CSUM) ||
205 			     !(type & (SKB_GSO_UDP))))
206 			goto out;
207 
208 		skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
209 
210 		segs = NULL;
211 		goto out;
212 	}
213 
214 	/* Do software UFO. Complete and fill in the UDP checksum as
215 	 * HW cannot do checksum of UDP packets sent as multiple
216 	 * IP fragments.
217 	 */
218 
219 	uh = udp_hdr(skb);
220 	iph = ip_hdr(skb);
221 
222 	uh->check = 0;
223 	csum = skb_checksum(skb, 0, skb->len, 0);
224 	uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
225 	if (uh->check == 0)
226 		uh->check = CSUM_MANGLED_0;
227 
228 	skb->ip_summed = CHECKSUM_NONE;
229 
230 	/* If there is no outer header we can fake a checksum offload
231 	 * due to the fact that we have already done the checksum in
232 	 * software prior to segmenting the frame.
233 	 */
234 	if (!skb->encap_hdr_csum)
235 		features |= NETIF_F_HW_CSUM;
236 
237 	/* Fragment the skb. IP headers of the fragments are updated in
238 	 * inet_gso_segment()
239 	 */
240 	segs = skb_segment(skb, features);
241 out:
242 	return segs;
243 }
244 
245 struct sk_buff **udp_gro_receive(struct sk_buff **head, struct sk_buff *skb,
246 				 struct udphdr *uh, udp_lookup_t lookup)
247 {
248 	struct sk_buff *p, **pp = NULL;
249 	struct udphdr *uh2;
250 	unsigned int off = skb_gro_offset(skb);
251 	int flush = 1;
252 	struct sock *sk;
253 
254 	if (NAPI_GRO_CB(skb)->encap_mark ||
255 	    (skb->ip_summed != CHECKSUM_PARTIAL &&
256 	     NAPI_GRO_CB(skb)->csum_cnt == 0 &&
257 	     !NAPI_GRO_CB(skb)->csum_valid))
258 		goto out;
259 
260 	/* mark that this skb passed once through the tunnel gro layer */
261 	NAPI_GRO_CB(skb)->encap_mark = 1;
262 
263 	rcu_read_lock();
264 	sk = (*lookup)(skb, uh->source, uh->dest);
265 
266 	if (sk && udp_sk(sk)->gro_receive)
267 		goto unflush;
268 	goto out_unlock;
269 
270 unflush:
271 	flush = 0;
272 
273 	for (p = *head; p; p = p->next) {
274 		if (!NAPI_GRO_CB(p)->same_flow)
275 			continue;
276 
277 		uh2 = (struct udphdr   *)(p->data + off);
278 
279 		/* Match ports and either checksums are either both zero
280 		 * or nonzero.
281 		 */
282 		if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
283 		    (!uh->check ^ !uh2->check)) {
284 			NAPI_GRO_CB(p)->same_flow = 0;
285 			continue;
286 		}
287 	}
288 
289 	skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
290 	skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
291 	pp = udp_sk(sk)->gro_receive(sk, head, skb);
292 
293 out_unlock:
294 	rcu_read_unlock();
295 out:
296 	NAPI_GRO_CB(skb)->flush |= flush;
297 	return pp;
298 }
299 EXPORT_SYMBOL(udp_gro_receive);
300 
301 static struct sk_buff **udp4_gro_receive(struct sk_buff **head,
302 					 struct sk_buff *skb)
303 {
304 	struct udphdr *uh = udp_gro_udphdr(skb);
305 
306 	if (unlikely(!uh))
307 		goto flush;
308 
309 	/* Don't bother verifying checksum if we're going to flush anyway. */
310 	if (NAPI_GRO_CB(skb)->flush)
311 		goto skip;
312 
313 	if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
314 						 inet_gro_compute_pseudo))
315 		goto flush;
316 	else if (uh->check)
317 		skb_gro_checksum_try_convert(skb, IPPROTO_UDP, uh->check,
318 					     inet_gro_compute_pseudo);
319 skip:
320 	NAPI_GRO_CB(skb)->is_ipv6 = 0;
321 	return udp_gro_receive(head, skb, uh, udp4_lib_lookup_skb);
322 
323 flush:
324 	NAPI_GRO_CB(skb)->flush = 1;
325 	return NULL;
326 }
327 
328 int udp_gro_complete(struct sk_buff *skb, int nhoff,
329 		     udp_lookup_t lookup)
330 {
331 	__be16 newlen = htons(skb->len - nhoff);
332 	struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
333 	int err = -ENOSYS;
334 	struct sock *sk;
335 
336 	uh->len = newlen;
337 
338 	rcu_read_lock();
339 	sk = (*lookup)(skb, uh->source, uh->dest);
340 	if (sk && udp_sk(sk)->gro_complete)
341 		err = udp_sk(sk)->gro_complete(sk, skb,
342 				nhoff + sizeof(struct udphdr));
343 	rcu_read_unlock();
344 
345 	if (skb->remcsum_offload)
346 		skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
347 
348 	skb->encapsulation = 1;
349 	skb_set_inner_mac_header(skb, nhoff + sizeof(struct udphdr));
350 
351 	return err;
352 }
353 EXPORT_SYMBOL(udp_gro_complete);
354 
355 static int udp4_gro_complete(struct sk_buff *skb, int nhoff)
356 {
357 	const struct iphdr *iph = ip_hdr(skb);
358 	struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
359 
360 	if (uh->check) {
361 		skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL_CSUM;
362 		uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
363 					  iph->daddr, 0);
364 	} else {
365 		skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL;
366 	}
367 
368 	return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
369 }
370 
371 static const struct net_offload udpv4_offload = {
372 	.callbacks = {
373 		.gso_segment = udp4_ufo_fragment,
374 		.gro_receive  =	udp4_gro_receive,
375 		.gro_complete =	udp4_gro_complete,
376 	},
377 };
378 
379 int __init udpv4_offload_init(void)
380 {
381 	return inet_add_offload(&udpv4_offload, IPPROTO_UDP);
382 }
383