xref: /openbmc/linux/net/ipv4/tcp_offload.c (revision 55b24334)
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 	unsigned int oldlen;
64 	unsigned int mss;
65 	struct sk_buff *gso_skb = skb;
66 	__sum16 newcheck;
67 	bool ooo_okay, copy_destructor;
68 	__wsum delta;
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 = ~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 = (__force __wsum)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(csum_add(csum_unfold(th->check), 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 		int delta;
153 
154 		swap(gso_skb->sk, skb->sk);
155 		swap(gso_skb->destructor, skb->destructor);
156 		sum_truesize += skb->truesize;
157 		delta = sum_truesize - gso_skb->truesize;
158 		/* In some pathological cases, delta can be negative.
159 		 * We need to either use refcount_add() or refcount_sub_and_test()
160 		 */
161 		if (likely(delta >= 0))
162 			refcount_add(delta, &skb->sk->sk_wmem_alloc);
163 		else
164 			WARN_ON_ONCE(refcount_sub_and_test(-delta, &skb->sk->sk_wmem_alloc));
165 	}
166 
167 	delta = (__force __wsum)htonl(oldlen +
168 				      (skb_tail_pointer(skb) -
169 				       skb_transport_header(skb)) +
170 				      skb->data_len);
171 	th->check = ~csum_fold(csum_add(csum_unfold(th->check), delta));
172 	if (skb->ip_summed == CHECKSUM_PARTIAL)
173 		gso_reset_checksum(skb, ~th->check);
174 	else
175 		th->check = gso_make_checksum(skb, ~th->check);
176 out:
177 	return segs;
178 }
179 
180 struct sk_buff *tcp_gro_receive(struct list_head *head, struct sk_buff *skb)
181 {
182 	struct sk_buff *pp = NULL;
183 	struct sk_buff *p;
184 	struct tcphdr *th;
185 	struct tcphdr *th2;
186 	unsigned int len;
187 	unsigned int thlen;
188 	__be32 flags;
189 	unsigned int mss = 1;
190 	unsigned int hlen;
191 	unsigned int off;
192 	int flush = 1;
193 	int i;
194 
195 	off = skb_gro_offset(skb);
196 	hlen = off + sizeof(*th);
197 	th = skb_gro_header(skb, hlen, off);
198 	if (unlikely(!th))
199 		goto out;
200 
201 	thlen = th->doff * 4;
202 	if (thlen < sizeof(*th))
203 		goto out;
204 
205 	hlen = off + thlen;
206 	if (skb_gro_header_hard(skb, hlen)) {
207 		th = skb_gro_header_slow(skb, hlen, off);
208 		if (unlikely(!th))
209 			goto out;
210 	}
211 
212 	skb_gro_pull(skb, thlen);
213 
214 	len = skb_gro_len(skb);
215 	flags = tcp_flag_word(th);
216 
217 	list_for_each_entry(p, head, list) {
218 		if (!NAPI_GRO_CB(p)->same_flow)
219 			continue;
220 
221 		th2 = tcp_hdr(p);
222 
223 		if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
224 			NAPI_GRO_CB(p)->same_flow = 0;
225 			continue;
226 		}
227 
228 		goto found;
229 	}
230 	p = NULL;
231 	goto out_check_final;
232 
233 found:
234 	/* Include the IP ID check below from the inner most IP hdr */
235 	flush = NAPI_GRO_CB(p)->flush;
236 	flush |= (__force int)(flags & TCP_FLAG_CWR);
237 	flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
238 		  ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
239 	flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
240 	for (i = sizeof(*th); i < thlen; i += 4)
241 		flush |= *(u32 *)((u8 *)th + i) ^
242 			 *(u32 *)((u8 *)th2 + i);
243 
244 	/* When we receive our second frame we can made a decision on if we
245 	 * continue this flow as an atomic flow with a fixed ID or if we use
246 	 * an incrementing ID.
247 	 */
248 	if (NAPI_GRO_CB(p)->flush_id != 1 ||
249 	    NAPI_GRO_CB(p)->count != 1 ||
250 	    !NAPI_GRO_CB(p)->is_atomic)
251 		flush |= NAPI_GRO_CB(p)->flush_id;
252 	else
253 		NAPI_GRO_CB(p)->is_atomic = false;
254 
255 	mss = skb_shinfo(p)->gso_size;
256 
257 	/* If skb is a GRO packet, make sure its gso_size matches prior packet mss.
258 	 * If it is a single frame, do not aggregate it if its length
259 	 * is bigger than our mss.
260 	 */
261 	if (unlikely(skb_is_gso(skb)))
262 		flush |= (mss != skb_shinfo(skb)->gso_size);
263 	else
264 		flush |= (len - 1) >= mss;
265 
266 	flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
267 #ifdef CONFIG_TLS_DEVICE
268 	flush |= p->decrypted ^ skb->decrypted;
269 #endif
270 
271 	if (flush || skb_gro_receive(p, skb)) {
272 		mss = 1;
273 		goto out_check_final;
274 	}
275 
276 	tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
277 
278 out_check_final:
279 	/* Force a flush if last segment is smaller than mss. */
280 	if (unlikely(skb_is_gso(skb)))
281 		flush = len != NAPI_GRO_CB(skb)->count * skb_shinfo(skb)->gso_size;
282 	else
283 		flush = len < mss;
284 
285 	flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
286 					TCP_FLAG_RST | TCP_FLAG_SYN |
287 					TCP_FLAG_FIN));
288 
289 	if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
290 		pp = p;
291 
292 out:
293 	NAPI_GRO_CB(skb)->flush |= (flush != 0);
294 
295 	return pp;
296 }
297 
298 void tcp_gro_complete(struct sk_buff *skb)
299 {
300 	struct tcphdr *th = tcp_hdr(skb);
301 
302 	skb->csum_start = (unsigned char *)th - skb->head;
303 	skb->csum_offset = offsetof(struct tcphdr, check);
304 	skb->ip_summed = CHECKSUM_PARTIAL;
305 
306 	skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
307 
308 	if (th->cwr)
309 		skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
310 
311 	if (skb->encapsulation)
312 		skb->inner_transport_header = skb->transport_header;
313 }
314 EXPORT_SYMBOL(tcp_gro_complete);
315 
316 INDIRECT_CALLABLE_SCOPE
317 struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb)
318 {
319 	/* Don't bother verifying checksum if we're going to flush anyway. */
320 	if (!NAPI_GRO_CB(skb)->flush &&
321 	    skb_gro_checksum_validate(skb, IPPROTO_TCP,
322 				      inet_gro_compute_pseudo)) {
323 		NAPI_GRO_CB(skb)->flush = 1;
324 		return NULL;
325 	}
326 
327 	return tcp_gro_receive(head, skb);
328 }
329 
330 INDIRECT_CALLABLE_SCOPE int tcp4_gro_complete(struct sk_buff *skb, int thoff)
331 {
332 	const struct iphdr *iph = ip_hdr(skb);
333 	struct tcphdr *th = tcp_hdr(skb);
334 
335 	th->check = ~tcp_v4_check(skb->len - thoff, iph->saddr,
336 				  iph->daddr, 0);
337 	skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
338 
339 	if (NAPI_GRO_CB(skb)->is_atomic)
340 		skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_FIXEDID;
341 
342 	tcp_gro_complete(skb);
343 	return 0;
344 }
345 
346 static const struct net_offload tcpv4_offload = {
347 	.callbacks = {
348 		.gso_segment	=	tcp4_gso_segment,
349 		.gro_receive	=	tcp4_gro_receive,
350 		.gro_complete	=	tcp4_gro_complete,
351 	},
352 };
353 
354 int __init tcpv4_offload_init(void)
355 {
356 	return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);
357 }
358