xref: /linux/net/ipv4/tcp_offload.c (revision 1cac38910ecb881b09f61f57545a771bbe57ba68)
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/gso.h>
13 #include <net/tcp.h>
14 #include <net/protocol.h>
15 
tcp_gso_tstamp(struct sk_buff * skb,struct sk_buff * gso_skb,unsigned int seq,unsigned int mss)16 static void tcp_gso_tstamp(struct sk_buff *skb, struct sk_buff *gso_skb,
17 			   unsigned int seq, unsigned int mss)
18 {
19 	u32 flags = skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP;
20 	u32 ts_seq = skb_shinfo(gso_skb)->tskey;
21 
22 	while (skb) {
23 		if (before(ts_seq, seq + mss)) {
24 			skb_shinfo(skb)->tx_flags |= flags;
25 			skb_shinfo(skb)->tskey = ts_seq;
26 			return;
27 		}
28 
29 		skb = skb->next;
30 		seq += mss;
31 	}
32 }
33 
__tcpv4_gso_segment_csum(struct sk_buff * seg,__be32 * oldip,__be32 newip,__be16 * oldport,__be16 newport)34 static void __tcpv4_gso_segment_csum(struct sk_buff *seg,
35 				     __be32 *oldip, __be32 newip,
36 				     __be16 *oldport, __be16 newport)
37 {
38 	struct tcphdr *th;
39 	struct iphdr *iph;
40 
41 	if (*oldip == newip && *oldport == newport)
42 		return;
43 
44 	th = tcp_hdr(seg);
45 	iph = ip_hdr(seg);
46 
47 	inet_proto_csum_replace4(&th->check, seg, *oldip, newip, true);
48 	inet_proto_csum_replace2(&th->check, seg, *oldport, newport, false);
49 	*oldport = newport;
50 
51 	csum_replace4(&iph->check, *oldip, newip);
52 	*oldip = newip;
53 }
54 
__tcpv4_gso_segment_list_csum(struct sk_buff * segs)55 static struct sk_buff *__tcpv4_gso_segment_list_csum(struct sk_buff *segs)
56 {
57 	const struct tcphdr *th;
58 	const struct iphdr *iph;
59 	struct sk_buff *seg;
60 	struct tcphdr *th2;
61 	struct iphdr *iph2;
62 
63 	seg = segs;
64 	th = tcp_hdr(seg);
65 	iph = ip_hdr(seg);
66 	th2 = tcp_hdr(seg->next);
67 	iph2 = ip_hdr(seg->next);
68 
69 	if (!(*(const u32 *)&th->source ^ *(const u32 *)&th2->source) &&
70 	    iph->daddr == iph2->daddr && iph->saddr == iph2->saddr)
71 		return segs;
72 
73 	while ((seg = seg->next)) {
74 		th2 = tcp_hdr(seg);
75 		iph2 = ip_hdr(seg);
76 
77 		__tcpv4_gso_segment_csum(seg,
78 					 &iph2->saddr, iph->saddr,
79 					 &th2->source, th->source);
80 		__tcpv4_gso_segment_csum(seg,
81 					 &iph2->daddr, iph->daddr,
82 					 &th2->dest, th->dest);
83 	}
84 
85 	return segs;
86 }
87 
__tcp4_gso_segment_list(struct sk_buff * skb,netdev_features_t features)88 static struct sk_buff *__tcp4_gso_segment_list(struct sk_buff *skb,
89 					      netdev_features_t features)
90 {
91 	skb = skb_segment_list(skb, features, skb_mac_header_len(skb));
92 	if (IS_ERR(skb))
93 		return skb;
94 
95 	return __tcpv4_gso_segment_list_csum(skb);
96 }
97 
tcp4_gso_segment(struct sk_buff * skb,netdev_features_t features)98 static struct sk_buff *tcp4_gso_segment(struct sk_buff *skb,
99 					netdev_features_t features)
100 {
101 	if (!(skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4))
102 		return ERR_PTR(-EINVAL);
103 
104 	if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
105 		return ERR_PTR(-EINVAL);
106 
107 	if (skb_shinfo(skb)->gso_type & SKB_GSO_FRAGLIST) {
108 		struct tcphdr *th = tcp_hdr(skb);
109 
110 		if ((skb_pagelen(skb) - th->doff * 4 == skb_shinfo(skb)->gso_size) &&
111 		    !(skb_shinfo(skb)->gso_type & SKB_GSO_DODGY))
112 			return __tcp4_gso_segment_list(skb, features);
113 
114 		skb->ip_summed = CHECKSUM_NONE;
115 	}
116 
117 	if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
118 		const struct iphdr *iph = ip_hdr(skb);
119 		struct tcphdr *th = tcp_hdr(skb);
120 
121 		/* Set up checksum pseudo header, usually expect stack to
122 		 * have done this already.
123 		 */
124 
125 		th->check = 0;
126 		skb->ip_summed = CHECKSUM_PARTIAL;
127 		__tcp_v4_send_check(skb, iph->saddr, iph->daddr);
128 	}
129 
130 	return tcp_gso_segment(skb, features);
131 }
132 
tcp_gso_segment(struct sk_buff * skb,netdev_features_t features)133 struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
134 				netdev_features_t features)
135 {
136 	struct sk_buff *segs = ERR_PTR(-EINVAL);
137 	unsigned int sum_truesize = 0;
138 	struct tcphdr *th;
139 	unsigned int thlen;
140 	unsigned int seq;
141 	unsigned int oldlen;
142 	unsigned int mss;
143 	struct sk_buff *gso_skb = skb;
144 	__sum16 newcheck;
145 	bool ooo_okay, copy_destructor;
146 	bool ecn_cwr_mask;
147 	__wsum delta;
148 
149 	th = tcp_hdr(skb);
150 	thlen = th->doff * 4;
151 	if (thlen < sizeof(*th))
152 		goto out;
153 
154 	if (unlikely(skb_checksum_start(skb) != skb_transport_header(skb)))
155 		goto out;
156 
157 	if (!pskb_may_pull(skb, thlen))
158 		goto out;
159 
160 	oldlen = ~skb->len;
161 	__skb_pull(skb, thlen);
162 
163 	mss = skb_shinfo(skb)->gso_size;
164 	if (unlikely(skb->len <= mss))
165 		goto out;
166 
167 	if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
168 		/* Packet is from an untrusted source, reset gso_segs. */
169 
170 		skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
171 
172 		segs = NULL;
173 		goto out;
174 	}
175 
176 	copy_destructor = gso_skb->destructor == tcp_wfree;
177 	ooo_okay = gso_skb->ooo_okay;
178 	/* All segments but the first should have ooo_okay cleared */
179 	skb->ooo_okay = 0;
180 
181 	segs = skb_segment(skb, features);
182 	if (IS_ERR(segs))
183 		goto out;
184 
185 	/* Only first segment might have ooo_okay set */
186 	segs->ooo_okay = ooo_okay;
187 
188 	/* GSO partial and frag_list segmentation only requires splitting
189 	 * the frame into an MSS multiple and possibly a remainder, both
190 	 * cases return a GSO skb. So update the mss now.
191 	 */
192 	if (skb_is_gso(segs))
193 		mss *= skb_shinfo(segs)->gso_segs;
194 
195 	delta = (__force __wsum)htonl(oldlen + thlen + mss);
196 
197 	skb = segs;
198 	th = tcp_hdr(skb);
199 	seq = ntohl(th->seq);
200 
201 	if (unlikely(skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP))
202 		tcp_gso_tstamp(segs, gso_skb, seq, mss);
203 
204 	newcheck = ~csum_fold(csum_add(csum_unfold(th->check), delta));
205 
206 	ecn_cwr_mask = !!(skb_shinfo(gso_skb)->gso_type & SKB_GSO_TCP_ACCECN);
207 
208 	while (skb->next) {
209 		th->fin = th->psh = 0;
210 		th->check = newcheck;
211 
212 		if (skb->ip_summed == CHECKSUM_PARTIAL)
213 			gso_reset_checksum(skb, ~th->check);
214 		else
215 			th->check = gso_make_checksum(skb, ~th->check);
216 
217 		seq += mss;
218 		if (copy_destructor) {
219 			skb->destructor = gso_skb->destructor;
220 			skb->sk = gso_skb->sk;
221 			sum_truesize += skb->truesize;
222 		}
223 		skb = skb->next;
224 		th = tcp_hdr(skb);
225 
226 		th->seq = htonl(seq);
227 
228 		th->cwr &= ecn_cwr_mask;
229 	}
230 
231 	/* Following permits TCP Small Queues to work well with GSO :
232 	 * The callback to TCP stack will be called at the time last frag
233 	 * is freed at TX completion, and not right now when gso_skb
234 	 * is freed by GSO engine
235 	 */
236 	if (copy_destructor) {
237 		int delta;
238 
239 		swap(gso_skb->sk, skb->sk);
240 		swap(gso_skb->destructor, skb->destructor);
241 		sum_truesize += skb->truesize;
242 		delta = sum_truesize - gso_skb->truesize;
243 		/* In some pathological cases, delta can be negative.
244 		 * We need to either use refcount_add() or refcount_sub_and_test()
245 		 */
246 		if (likely(delta >= 0))
247 			refcount_add(delta, &skb->sk->sk_wmem_alloc);
248 		else
249 			WARN_ON_ONCE(refcount_sub_and_test(-delta, &skb->sk->sk_wmem_alloc));
250 	}
251 
252 	delta = (__force __wsum)htonl(oldlen +
253 				      (skb_tail_pointer(skb) -
254 				       skb_transport_header(skb)) +
255 				      skb->data_len);
256 	th->check = ~csum_fold(csum_add(csum_unfold(th->check), delta));
257 	if (skb->ip_summed == CHECKSUM_PARTIAL)
258 		gso_reset_checksum(skb, ~th->check);
259 	else
260 		th->check = gso_make_checksum(skb, ~th->check);
261 out:
262 	return segs;
263 }
264 
tcp_gro_lookup(struct list_head * head,struct tcphdr * th)265 struct sk_buff *tcp_gro_lookup(struct list_head *head, struct tcphdr *th)
266 {
267 	struct tcphdr *th2;
268 	struct sk_buff *p;
269 
270 	list_for_each_entry(p, head, list) {
271 		if (!NAPI_GRO_CB(p)->same_flow)
272 			continue;
273 
274 		th2 = tcp_hdr(p);
275 		if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
276 			NAPI_GRO_CB(p)->same_flow = 0;
277 			continue;
278 		}
279 
280 		return p;
281 	}
282 
283 	return NULL;
284 }
285 
tcp_gro_receive(struct list_head * head,struct sk_buff * skb,struct tcphdr * th)286 struct sk_buff *tcp_gro_receive(struct list_head *head, struct sk_buff *skb,
287 				struct tcphdr *th)
288 {
289 	unsigned int thlen = th->doff * 4;
290 	struct sk_buff *pp = NULL;
291 	struct sk_buff *p;
292 	struct tcphdr *th2;
293 	unsigned int len;
294 	__be32 flags;
295 	unsigned int mss = 1;
296 	int flush = 1;
297 	int i;
298 
299 	len = skb_gro_len(skb);
300 	flags = tcp_flag_word(th);
301 
302 	p = tcp_gro_lookup(head, th);
303 	if (!p)
304 		goto out_check_final;
305 
306 	th2 = tcp_hdr(p);
307 	flush = (__force int)(flags & TCP_FLAG_CWR);
308 	flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
309 		  ~(TCP_FLAG_FIN | TCP_FLAG_PSH));
310 	flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
311 	for (i = sizeof(*th); i < thlen; i += 4)
312 		flush |= *(u32 *)((u8 *)th + i) ^
313 			 *(u32 *)((u8 *)th2 + i);
314 
315 	flush |= gro_receive_network_flush(th, th2, p);
316 
317 	mss = skb_shinfo(p)->gso_size;
318 
319 	/* If skb is a GRO packet, make sure its gso_size matches prior packet mss.
320 	 * If it is a single frame, do not aggregate it if its length
321 	 * is bigger than our mss.
322 	 */
323 	if (unlikely(skb_is_gso(skb)))
324 		flush |= (mss != skb_shinfo(skb)->gso_size);
325 	else
326 		flush |= (len - 1) >= mss;
327 
328 	flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
329 	flush |= skb_cmp_decrypted(p, skb);
330 
331 	if (unlikely(NAPI_GRO_CB(p)->is_flist)) {
332 		flush |= (__force int)(flags ^ tcp_flag_word(th2));
333 		flush |= skb->ip_summed != p->ip_summed;
334 		flush |= skb->csum_level != p->csum_level;
335 		flush |= NAPI_GRO_CB(p)->count >= 64;
336 		skb_set_network_header(skb, skb_gro_receive_network_offset(skb));
337 
338 		if (flush || skb_gro_receive_list(p, skb))
339 			mss = 1;
340 
341 		goto out_check_final;
342 	}
343 
344 	if (flush || skb_gro_receive(p, skb)) {
345 		mss = 1;
346 		goto out_check_final;
347 	}
348 
349 	tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
350 
351 out_check_final:
352 	/* Force a flush if last segment is smaller than mss. */
353 	if (unlikely(skb_is_gso(skb)))
354 		flush = len != NAPI_GRO_CB(skb)->count * skb_shinfo(skb)->gso_size;
355 	else
356 		flush = len < mss;
357 
358 	flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
359 					TCP_FLAG_RST | TCP_FLAG_SYN |
360 					TCP_FLAG_FIN));
361 
362 	if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
363 		pp = p;
364 
365 	NAPI_GRO_CB(skb)->flush |= (flush != 0);
366 
367 	return pp;
368 }
369 
tcp_gro_complete(struct sk_buff * skb)370 void tcp_gro_complete(struct sk_buff *skb)
371 {
372 	struct tcphdr *th = tcp_hdr(skb);
373 	struct skb_shared_info *shinfo;
374 
375 	if (skb->encapsulation)
376 		skb->inner_transport_header = skb->transport_header;
377 
378 	skb->csum_start = (unsigned char *)th - skb->head;
379 	skb->csum_offset = offsetof(struct tcphdr, check);
380 	skb->ip_summed = CHECKSUM_PARTIAL;
381 
382 	shinfo = skb_shinfo(skb);
383 	shinfo->gso_segs = NAPI_GRO_CB(skb)->count;
384 
385 	if (th->cwr)
386 		shinfo->gso_type |= SKB_GSO_TCP_ACCECN;
387 }
388 EXPORT_SYMBOL(tcp_gro_complete);
389 
tcp4_check_fraglist_gro(struct list_head * head,struct sk_buff * skb,struct tcphdr * th)390 static void tcp4_check_fraglist_gro(struct list_head *head, struct sk_buff *skb,
391 				    struct tcphdr *th)
392 {
393 	const struct iphdr *iph;
394 	struct sk_buff *p;
395 	struct sock *sk;
396 	struct net *net;
397 	int iif, sdif;
398 
399 	if (likely(!(skb->dev->features & NETIF_F_GRO_FRAGLIST)))
400 		return;
401 
402 	p = tcp_gro_lookup(head, th);
403 	if (p) {
404 		NAPI_GRO_CB(skb)->is_flist = NAPI_GRO_CB(p)->is_flist;
405 		return;
406 	}
407 
408 	inet_get_iif_sdif(skb, &iif, &sdif);
409 	iph = skb_gro_network_header(skb);
410 	net = dev_net_rcu(skb->dev);
411 	sk = __inet_lookup_established(net, iph->saddr, th->source,
412 				       iph->daddr, ntohs(th->dest),
413 				       iif, sdif);
414 	NAPI_GRO_CB(skb)->is_flist = !sk;
415 	if (sk)
416 		sock_gen_put(sk);
417 }
418 
419 INDIRECT_CALLABLE_SCOPE
tcp4_gro_receive(struct list_head * head,struct sk_buff * skb)420 struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb)
421 {
422 	struct tcphdr *th;
423 
424 	/* Don't bother verifying checksum if we're going to flush anyway. */
425 	if (!NAPI_GRO_CB(skb)->flush &&
426 	    skb_gro_checksum_validate(skb, IPPROTO_TCP,
427 				      inet_gro_compute_pseudo))
428 		goto flush;
429 
430 	th = tcp_gro_pull_header(skb);
431 	if (!th)
432 		goto flush;
433 
434 	tcp4_check_fraglist_gro(head, skb, th);
435 
436 	return tcp_gro_receive(head, skb, th);
437 
438 flush:
439 	NAPI_GRO_CB(skb)->flush = 1;
440 	return NULL;
441 }
442 
tcp4_gro_complete(struct sk_buff * skb,int thoff)443 INDIRECT_CALLABLE_SCOPE int tcp4_gro_complete(struct sk_buff *skb, int thoff)
444 {
445 	const u16 offset = NAPI_GRO_CB(skb)->network_offsets[skb->encapsulation];
446 	const struct iphdr *iph = (struct iphdr *)(skb->data + offset);
447 	struct tcphdr *th = tcp_hdr(skb);
448 
449 	if (unlikely(NAPI_GRO_CB(skb)->is_flist)) {
450 		skb_shinfo(skb)->gso_type |= SKB_GSO_FRAGLIST | SKB_GSO_TCPV4;
451 		skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
452 
453 		__skb_incr_checksum_unnecessary(skb);
454 
455 		return 0;
456 	}
457 
458 	th->check = ~tcp_v4_check(skb->len - thoff, iph->saddr,
459 				  iph->daddr, 0);
460 
461 	BUILD_BUG_ON(SKB_GSO_TCP_FIXEDID << 1 != SKB_GSO_TCP_FIXEDID_INNER);
462 	skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4 |
463 			(NAPI_GRO_CB(skb)->ip_fixedid * SKB_GSO_TCP_FIXEDID);
464 
465 	tcp_gro_complete(skb);
466 	return 0;
467 }
468 
tcpv4_offload_init(void)469 int __init tcpv4_offload_init(void)
470 {
471 	net_hotdata.tcpv4_offload = (struct net_offload) {
472 		.callbacks = {
473 			.gso_segment	=	tcp4_gso_segment,
474 			.gro_receive	=	tcp4_gro_receive,
475 			.gro_complete	=	tcp4_gro_complete,
476 		},
477 	};
478 	return inet_add_offload(&net_hotdata.tcpv4_offload, IPPROTO_TCP);
479 }
480