xref: /linux/net/ipv4/tcp_offload.c (revision 80d443e8876602be2c130f79c4de81e12e2a700d)
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  *	TCPv4 GSO/GRO support
11  */
12 
13 #include <linux/skbuff.h>
14 #include <net/tcp.h>
15 #include <net/protocol.h>
16 
17 static void tcp_gso_tstamp(struct sk_buff *skb, unsigned int ts_seq,
18 			   unsigned int seq, unsigned int mss)
19 {
20 	while (skb) {
21 		if (before(ts_seq, seq + mss)) {
22 			skb_shinfo(skb)->tx_flags |= SKBTX_SW_TSTAMP;
23 			skb_shinfo(skb)->tskey = ts_seq;
24 			return;
25 		}
26 
27 		skb = skb->next;
28 		seq += mss;
29 	}
30 }
31 
32 static struct sk_buff *tcp4_gso_segment(struct sk_buff *skb,
33 					netdev_features_t features)
34 {
35 	if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
36 		return ERR_PTR(-EINVAL);
37 
38 	if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
39 		const struct iphdr *iph = ip_hdr(skb);
40 		struct tcphdr *th = tcp_hdr(skb);
41 
42 		/* Set up checksum pseudo header, usually expect stack to
43 		 * have done this already.
44 		 */
45 
46 		th->check = 0;
47 		skb->ip_summed = CHECKSUM_PARTIAL;
48 		__tcp_v4_send_check(skb, iph->saddr, iph->daddr);
49 	}
50 
51 	return tcp_gso_segment(skb, features);
52 }
53 
54 struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
55 				netdev_features_t features)
56 {
57 	struct sk_buff *segs = ERR_PTR(-EINVAL);
58 	unsigned int sum_truesize = 0;
59 	struct tcphdr *th;
60 	unsigned int thlen;
61 	unsigned int seq;
62 	__be32 delta;
63 	unsigned int oldlen;
64 	unsigned int mss;
65 	struct sk_buff *gso_skb = skb;
66 	__sum16 newcheck;
67 	bool ooo_okay, copy_destructor;
68 
69 	th = tcp_hdr(skb);
70 	thlen = th->doff * 4;
71 	if (thlen < sizeof(*th))
72 		goto out;
73 
74 	if (!pskb_may_pull(skb, thlen))
75 		goto out;
76 
77 	oldlen = (u16)~skb->len;
78 	__skb_pull(skb, thlen);
79 
80 	mss = skb_shinfo(skb)->gso_size;
81 	if (unlikely(skb->len <= mss))
82 		goto out;
83 
84 	if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
85 		/* Packet is from an untrusted source, reset gso_segs. */
86 
87 		skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
88 
89 		segs = NULL;
90 		goto out;
91 	}
92 
93 	copy_destructor = gso_skb->destructor == tcp_wfree;
94 	ooo_okay = gso_skb->ooo_okay;
95 	/* All segments but the first should have ooo_okay cleared */
96 	skb->ooo_okay = 0;
97 
98 	segs = skb_segment(skb, features);
99 	if (IS_ERR(segs))
100 		goto out;
101 
102 	/* Only first segment might have ooo_okay set */
103 	segs->ooo_okay = ooo_okay;
104 
105 	/* GSO partial and frag_list segmentation only requires splitting
106 	 * the frame into an MSS multiple and possibly a remainder, both
107 	 * cases return a GSO skb. So update the mss now.
108 	 */
109 	if (skb_is_gso(segs))
110 		mss *= skb_shinfo(segs)->gso_segs;
111 
112 	delta = htonl(oldlen + (thlen + mss));
113 
114 	skb = segs;
115 	th = tcp_hdr(skb);
116 	seq = ntohl(th->seq);
117 
118 	if (unlikely(skb_shinfo(gso_skb)->tx_flags & SKBTX_SW_TSTAMP))
119 		tcp_gso_tstamp(segs, skb_shinfo(gso_skb)->tskey, seq, mss);
120 
121 	newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
122 					       (__force u32)delta));
123 
124 	while (skb->next) {
125 		th->fin = th->psh = 0;
126 		th->check = newcheck;
127 
128 		if (skb->ip_summed == CHECKSUM_PARTIAL)
129 			gso_reset_checksum(skb, ~th->check);
130 		else
131 			th->check = gso_make_checksum(skb, ~th->check);
132 
133 		seq += mss;
134 		if (copy_destructor) {
135 			skb->destructor = gso_skb->destructor;
136 			skb->sk = gso_skb->sk;
137 			sum_truesize += skb->truesize;
138 		}
139 		skb = skb->next;
140 		th = tcp_hdr(skb);
141 
142 		th->seq = htonl(seq);
143 		th->cwr = 0;
144 	}
145 
146 	/* Following permits TCP Small Queues to work well with GSO :
147 	 * The callback to TCP stack will be called at the time last frag
148 	 * is freed at TX completion, and not right now when gso_skb
149 	 * is freed by GSO engine
150 	 */
151 	if (copy_destructor) {
152 		swap(gso_skb->sk, skb->sk);
153 		swap(gso_skb->destructor, skb->destructor);
154 		sum_truesize += skb->truesize;
155 		atomic_add(sum_truesize - gso_skb->truesize,
156 			   &skb->sk->sk_wmem_alloc);
157 	}
158 
159 	delta = htonl(oldlen + (skb_tail_pointer(skb) -
160 				skb_transport_header(skb)) +
161 		      skb->data_len);
162 	th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
163 				(__force u32)delta));
164 	if (skb->ip_summed == CHECKSUM_PARTIAL)
165 		gso_reset_checksum(skb, ~th->check);
166 	else
167 		th->check = gso_make_checksum(skb, ~th->check);
168 out:
169 	return segs;
170 }
171 
172 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
173 {
174 	struct sk_buff **pp = NULL;
175 	struct sk_buff *p;
176 	struct tcphdr *th;
177 	struct tcphdr *th2;
178 	unsigned int len;
179 	unsigned int thlen;
180 	__be32 flags;
181 	unsigned int mss = 1;
182 	unsigned int hlen;
183 	unsigned int off;
184 	int flush = 1;
185 	int i;
186 
187 	off = skb_gro_offset(skb);
188 	hlen = off + sizeof(*th);
189 	th = skb_gro_header_fast(skb, off);
190 	if (skb_gro_header_hard(skb, hlen)) {
191 		th = skb_gro_header_slow(skb, hlen, off);
192 		if (unlikely(!th))
193 			goto out;
194 	}
195 
196 	thlen = th->doff * 4;
197 	if (thlen < sizeof(*th))
198 		goto out;
199 
200 	hlen = off + thlen;
201 	if (skb_gro_header_hard(skb, hlen)) {
202 		th = skb_gro_header_slow(skb, hlen, off);
203 		if (unlikely(!th))
204 			goto out;
205 	}
206 
207 	skb_gro_pull(skb, thlen);
208 
209 	len = skb_gro_len(skb);
210 	flags = tcp_flag_word(th);
211 
212 	for (; (p = *head); head = &p->next) {
213 		if (!NAPI_GRO_CB(p)->same_flow)
214 			continue;
215 
216 		th2 = tcp_hdr(p);
217 
218 		if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
219 			NAPI_GRO_CB(p)->same_flow = 0;
220 			continue;
221 		}
222 
223 		goto found;
224 	}
225 
226 	goto out_check_final;
227 
228 found:
229 	/* Include the IP ID check below from the inner most IP hdr */
230 	flush = NAPI_GRO_CB(p)->flush;
231 	flush |= (__force int)(flags & TCP_FLAG_CWR);
232 	flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
233 		  ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
234 	flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
235 	for (i = sizeof(*th); i < thlen; i += 4)
236 		flush |= *(u32 *)((u8 *)th + i) ^
237 			 *(u32 *)((u8 *)th2 + i);
238 
239 	/* When we receive our second frame we can made a decision on if we
240 	 * continue this flow as an atomic flow with a fixed ID or if we use
241 	 * an incrementing ID.
242 	 */
243 	if (NAPI_GRO_CB(p)->flush_id != 1 ||
244 	    NAPI_GRO_CB(p)->count != 1 ||
245 	    !NAPI_GRO_CB(p)->is_atomic)
246 		flush |= NAPI_GRO_CB(p)->flush_id;
247 	else
248 		NAPI_GRO_CB(p)->is_atomic = false;
249 
250 	mss = skb_shinfo(p)->gso_size;
251 
252 	flush |= (len - 1) >= mss;
253 	flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
254 
255 	if (flush || skb_gro_receive(head, skb)) {
256 		mss = 1;
257 		goto out_check_final;
258 	}
259 
260 	p = *head;
261 	th2 = tcp_hdr(p);
262 	tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
263 
264 out_check_final:
265 	flush = len < mss;
266 	flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
267 					TCP_FLAG_RST | TCP_FLAG_SYN |
268 					TCP_FLAG_FIN));
269 
270 	if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
271 		pp = head;
272 
273 out:
274 	NAPI_GRO_CB(skb)->flush |= (flush != 0);
275 
276 	return pp;
277 }
278 
279 int tcp_gro_complete(struct sk_buff *skb)
280 {
281 	struct tcphdr *th = tcp_hdr(skb);
282 
283 	skb->csum_start = (unsigned char *)th - skb->head;
284 	skb->csum_offset = offsetof(struct tcphdr, check);
285 	skb->ip_summed = CHECKSUM_PARTIAL;
286 
287 	skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
288 
289 	if (th->cwr)
290 		skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
291 
292 	return 0;
293 }
294 EXPORT_SYMBOL(tcp_gro_complete);
295 
296 static struct sk_buff **tcp4_gro_receive(struct sk_buff **head, struct sk_buff *skb)
297 {
298 	/* Don't bother verifying checksum if we're going to flush anyway. */
299 	if (!NAPI_GRO_CB(skb)->flush &&
300 	    skb_gro_checksum_validate(skb, IPPROTO_TCP,
301 				      inet_gro_compute_pseudo)) {
302 		NAPI_GRO_CB(skb)->flush = 1;
303 		return NULL;
304 	}
305 
306 	return tcp_gro_receive(head, skb);
307 }
308 
309 static int tcp4_gro_complete(struct sk_buff *skb, int thoff)
310 {
311 	const struct iphdr *iph = ip_hdr(skb);
312 	struct tcphdr *th = tcp_hdr(skb);
313 
314 	th->check = ~tcp_v4_check(skb->len - thoff, iph->saddr,
315 				  iph->daddr, 0);
316 	skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
317 
318 	if (NAPI_GRO_CB(skb)->is_atomic)
319 		skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_FIXEDID;
320 
321 	return tcp_gro_complete(skb);
322 }
323 
324 static const struct net_offload tcpv4_offload = {
325 	.callbacks = {
326 		.gso_segment	=	tcp4_gso_segment,
327 		.gro_receive	=	tcp4_gro_receive,
328 		.gro_complete	=	tcp4_gro_complete,
329 	},
330 };
331 
332 int __init tcpv4_offload_init(void)
333 {
334 	return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);
335 }
336