xref: /linux/net/ipv4/tcp_offload.c (revision cc8c418b4fc09ed58ddd27b8e90ec797e9ca1e67)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  *	IPV4 GSO/GRO offload support
4  *	Linux INET implementation
5  *
6  *	TCPv4 GSO/GRO support
7  */
8 
9 #include <linux/indirect_call_wrapper.h>
10 #include <linux/skbuff.h>
11 #include <net/gro.h>
12 #include <net/tcp.h>
13 #include <net/protocol.h>
14 
15 static void tcp_gso_tstamp(struct sk_buff *skb, unsigned int ts_seq,
16 			   unsigned int seq, unsigned int mss)
17 {
18 	while (skb) {
19 		if (before(ts_seq, seq + mss)) {
20 			skb_shinfo(skb)->tx_flags |= SKBTX_SW_TSTAMP;
21 			skb_shinfo(skb)->tskey = ts_seq;
22 			return;
23 		}
24 
25 		skb = skb->next;
26 		seq += mss;
27 	}
28 }
29 
30 static struct sk_buff *tcp4_gso_segment(struct sk_buff *skb,
31 					netdev_features_t features)
32 {
33 	if (!(skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4))
34 		return ERR_PTR(-EINVAL);
35 
36 	if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
37 		return ERR_PTR(-EINVAL);
38 
39 	if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
40 		const struct iphdr *iph = ip_hdr(skb);
41 		struct tcphdr *th = tcp_hdr(skb);
42 
43 		/* Set up checksum pseudo header, usually expect stack to
44 		 * have done this already.
45 		 */
46 
47 		th->check = 0;
48 		skb->ip_summed = CHECKSUM_PARTIAL;
49 		__tcp_v4_send_check(skb, iph->saddr, iph->daddr);
50 	}
51 
52 	return tcp_gso_segment(skb, features);
53 }
54 
55 struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
56 				netdev_features_t features)
57 {
58 	struct sk_buff *segs = ERR_PTR(-EINVAL);
59 	unsigned int sum_truesize = 0;
60 	struct tcphdr *th;
61 	unsigned int thlen;
62 	unsigned int seq;
63 	__be32 delta;
64 	unsigned int oldlen;
65 	unsigned int mss;
66 	struct sk_buff *gso_skb = skb;
67 	__sum16 newcheck;
68 	bool ooo_okay, copy_destructor;
69 
70 	th = tcp_hdr(skb);
71 	thlen = th->doff * 4;
72 	if (thlen < sizeof(*th))
73 		goto out;
74 
75 	if (!pskb_may_pull(skb, thlen))
76 		goto out;
77 
78 	oldlen = (u16)~skb->len;
79 	__skb_pull(skb, thlen);
80 
81 	mss = skb_shinfo(skb)->gso_size;
82 	if (unlikely(skb->len <= mss))
83 		goto out;
84 
85 	if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
86 		/* Packet is from an untrusted source, reset gso_segs. */
87 
88 		skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
89 
90 		segs = NULL;
91 		goto out;
92 	}
93 
94 	copy_destructor = gso_skb->destructor == tcp_wfree;
95 	ooo_okay = gso_skb->ooo_okay;
96 	/* All segments but the first should have ooo_okay cleared */
97 	skb->ooo_okay = 0;
98 
99 	segs = skb_segment(skb, features);
100 	if (IS_ERR(segs))
101 		goto out;
102 
103 	/* Only first segment might have ooo_okay set */
104 	segs->ooo_okay = ooo_okay;
105 
106 	/* GSO partial and frag_list segmentation only requires splitting
107 	 * the frame into an MSS multiple and possibly a remainder, both
108 	 * cases return a GSO skb. So update the mss now.
109 	 */
110 	if (skb_is_gso(segs))
111 		mss *= skb_shinfo(segs)->gso_segs;
112 
113 	delta = htonl(oldlen + (thlen + mss));
114 
115 	skb = segs;
116 	th = tcp_hdr(skb);
117 	seq = ntohl(th->seq);
118 
119 	if (unlikely(skb_shinfo(gso_skb)->tx_flags & SKBTX_SW_TSTAMP))
120 		tcp_gso_tstamp(segs, skb_shinfo(gso_skb)->tskey, seq, mss);
121 
122 	newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
123 					       (__force u32)delta));
124 
125 	while (skb->next) {
126 		th->fin = th->psh = 0;
127 		th->check = newcheck;
128 
129 		if (skb->ip_summed == CHECKSUM_PARTIAL)
130 			gso_reset_checksum(skb, ~th->check);
131 		else
132 			th->check = gso_make_checksum(skb, ~th->check);
133 
134 		seq += mss;
135 		if (copy_destructor) {
136 			skb->destructor = gso_skb->destructor;
137 			skb->sk = gso_skb->sk;
138 			sum_truesize += skb->truesize;
139 		}
140 		skb = skb->next;
141 		th = tcp_hdr(skb);
142 
143 		th->seq = htonl(seq);
144 		th->cwr = 0;
145 	}
146 
147 	/* Following permits TCP Small Queues to work well with GSO :
148 	 * The callback to TCP stack will be called at the time last frag
149 	 * is freed at TX completion, and not right now when gso_skb
150 	 * is freed by GSO engine
151 	 */
152 	if (copy_destructor) {
153 		int delta;
154 
155 		swap(gso_skb->sk, skb->sk);
156 		swap(gso_skb->destructor, skb->destructor);
157 		sum_truesize += skb->truesize;
158 		delta = sum_truesize - gso_skb->truesize;
159 		/* In some pathological cases, delta can be negative.
160 		 * We need to either use refcount_add() or refcount_sub_and_test()
161 		 */
162 		if (likely(delta >= 0))
163 			refcount_add(delta, &skb->sk->sk_wmem_alloc);
164 		else
165 			WARN_ON_ONCE(refcount_sub_and_test(-delta, &skb->sk->sk_wmem_alloc));
166 	}
167 
168 	delta = htonl(oldlen + (skb_tail_pointer(skb) -
169 				skb_transport_header(skb)) +
170 		      skb->data_len);
171 	th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
172 				(__force u32)delta));
173 	if (skb->ip_summed == CHECKSUM_PARTIAL)
174 		gso_reset_checksum(skb, ~th->check);
175 	else
176 		th->check = gso_make_checksum(skb, ~th->check);
177 out:
178 	return segs;
179 }
180 
181 struct sk_buff *tcp_gro_receive(struct list_head *head, struct sk_buff *skb)
182 {
183 	struct sk_buff *pp = NULL;
184 	struct sk_buff *p;
185 	struct tcphdr *th;
186 	struct tcphdr *th2;
187 	unsigned int len;
188 	unsigned int thlen;
189 	__be32 flags;
190 	unsigned int mss = 1;
191 	unsigned int hlen;
192 	unsigned int off;
193 	int flush = 1;
194 	int i;
195 
196 	off = skb_gro_offset(skb);
197 	hlen = off + sizeof(*th);
198 	th = skb_gro_header(skb, hlen, off);
199 	if (unlikely(!th))
200 		goto out;
201 
202 	thlen = th->doff * 4;
203 	if (thlen < sizeof(*th))
204 		goto out;
205 
206 	hlen = off + thlen;
207 	if (skb_gro_header_hard(skb, hlen)) {
208 		th = skb_gro_header_slow(skb, hlen, off);
209 		if (unlikely(!th))
210 			goto out;
211 	}
212 
213 	skb_gro_pull(skb, thlen);
214 
215 	len = skb_gro_len(skb);
216 	flags = tcp_flag_word(th);
217 
218 	list_for_each_entry(p, head, list) {
219 		if (!NAPI_GRO_CB(p)->same_flow)
220 			continue;
221 
222 		th2 = tcp_hdr(p);
223 
224 		if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
225 			NAPI_GRO_CB(p)->same_flow = 0;
226 			continue;
227 		}
228 
229 		goto found;
230 	}
231 	p = NULL;
232 	goto out_check_final;
233 
234 found:
235 	/* Include the IP ID check below from the inner most IP hdr */
236 	flush = NAPI_GRO_CB(p)->flush;
237 	flush |= (__force int)(flags & TCP_FLAG_CWR);
238 	flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
239 		  ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
240 	flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
241 	for (i = sizeof(*th); i < thlen; i += 4)
242 		flush |= *(u32 *)((u8 *)th + i) ^
243 			 *(u32 *)((u8 *)th2 + i);
244 
245 	/* When we receive our second frame we can made a decision on if we
246 	 * continue this flow as an atomic flow with a fixed ID or if we use
247 	 * an incrementing ID.
248 	 */
249 	if (NAPI_GRO_CB(p)->flush_id != 1 ||
250 	    NAPI_GRO_CB(p)->count != 1 ||
251 	    !NAPI_GRO_CB(p)->is_atomic)
252 		flush |= NAPI_GRO_CB(p)->flush_id;
253 	else
254 		NAPI_GRO_CB(p)->is_atomic = false;
255 
256 	mss = skb_shinfo(p)->gso_size;
257 
258 	/* If skb is a GRO packet, make sure its gso_size matches prior packet mss.
259 	 * If it is a single frame, do not aggregate it if its length
260 	 * is bigger than our mss.
261 	 */
262 	if (unlikely(skb_is_gso(skb)))
263 		flush |= (mss != skb_shinfo(skb)->gso_size);
264 	else
265 		flush |= (len - 1) >= mss;
266 
267 	flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
268 #ifdef CONFIG_TLS_DEVICE
269 	flush |= p->decrypted ^ skb->decrypted;
270 #endif
271 
272 	if (flush || skb_gro_receive(p, skb)) {
273 		mss = 1;
274 		goto out_check_final;
275 	}
276 
277 	tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
278 
279 out_check_final:
280 	/* Force a flush if last segment is smaller than mss. */
281 	if (unlikely(skb_is_gso(skb)))
282 		flush = len != NAPI_GRO_CB(skb)->count * skb_shinfo(skb)->gso_size;
283 	else
284 		flush = len < mss;
285 
286 	flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
287 					TCP_FLAG_RST | TCP_FLAG_SYN |
288 					TCP_FLAG_FIN));
289 
290 	if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
291 		pp = p;
292 
293 out:
294 	NAPI_GRO_CB(skb)->flush |= (flush != 0);
295 
296 	return pp;
297 }
298 
299 int tcp_gro_complete(struct sk_buff *skb)
300 {
301 	struct tcphdr *th = tcp_hdr(skb);
302 
303 	skb->csum_start = (unsigned char *)th - skb->head;
304 	skb->csum_offset = offsetof(struct tcphdr, check);
305 	skb->ip_summed = CHECKSUM_PARTIAL;
306 
307 	skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
308 
309 	if (th->cwr)
310 		skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
311 
312 	if (skb->encapsulation)
313 		skb->inner_transport_header = skb->transport_header;
314 
315 	return 0;
316 }
317 EXPORT_SYMBOL(tcp_gro_complete);
318 
319 INDIRECT_CALLABLE_SCOPE
320 struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb)
321 {
322 	/* Don't bother verifying checksum if we're going to flush anyway. */
323 	if (!NAPI_GRO_CB(skb)->flush &&
324 	    skb_gro_checksum_validate(skb, IPPROTO_TCP,
325 				      inet_gro_compute_pseudo)) {
326 		NAPI_GRO_CB(skb)->flush = 1;
327 		return NULL;
328 	}
329 
330 	return tcp_gro_receive(head, skb);
331 }
332 
333 INDIRECT_CALLABLE_SCOPE int tcp4_gro_complete(struct sk_buff *skb, int thoff)
334 {
335 	const struct iphdr *iph = ip_hdr(skb);
336 	struct tcphdr *th = tcp_hdr(skb);
337 
338 	th->check = ~tcp_v4_check(skb->len - thoff, iph->saddr,
339 				  iph->daddr, 0);
340 	skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
341 
342 	if (NAPI_GRO_CB(skb)->is_atomic)
343 		skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_FIXEDID;
344 
345 	return tcp_gro_complete(skb);
346 }
347 
348 static const struct net_offload tcpv4_offload = {
349 	.callbacks = {
350 		.gso_segment	=	tcp4_gso_segment,
351 		.gro_receive	=	tcp4_gro_receive,
352 		.gro_complete	=	tcp4_gro_complete,
353 	},
354 };
355 
356 int __init tcpv4_offload_init(void)
357 {
358 	return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);
359 }
360