?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge tag 'io_uring-5.11-2021-01-16' of git://git.kernel.dk/linux-block
[linux/fpc-iii.git] / net / ipv4 / udp_offload.c
blobff39e94781bfb825a6d24d7642424458550e6a12
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * IPV4 GSO/GRO offload support
4 * Linux INET implementation
5 *
6 * UDPv4 GSO support
7 */
8
9 #include <linux/skbuff.h>
10 #include <net/udp.h>
11 #include <net/protocol.h>
12 #include <net/inet_common.h>
13
14 static struct sk_buff *__skb_udp_tunnel_segment(struct sk_buff *skb,
15 netdev_features_t features,
16 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
17 netdev_features_t features),
18 __be16 new_protocol, bool is_ipv6)
19 {
20 int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
21 bool remcsum, need_csum, offload_csum, gso_partial;
22 struct sk_buff *segs = ERR_PTR(-EINVAL);
23 struct udphdr *uh = udp_hdr(skb);
24 u16 mac_offset = skb->mac_header;
25 __be16 protocol = skb->protocol;
26 u16 mac_len = skb->mac_len;
27 int udp_offset, outer_hlen;
28 __wsum partial;
29 bool need_ipsec;
30
31 if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
32 goto out;
33
34 /* Adjust partial header checksum to negate old length.
35 * We cannot rely on the value contained in uh->len as it is
36 * possible that the actual value exceeds the boundaries of the
37 * 16 bit length field due to the header being added outside of an
38 * IP or IPv6 frame that was already limited to 64K - 1.
39 */
40 if (skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)
41 partial = (__force __wsum)uh->len;
42 else
43 partial = (__force __wsum)htonl(skb->len);
44 partial = csum_sub(csum_unfold(uh->check), partial);
45
46 /* setup inner skb. */
47 skb->encapsulation = 0;
48 SKB_GSO_CB(skb)->encap_level = 0;
49 __skb_pull(skb, tnl_hlen);
50 skb_reset_mac_header(skb);
51 skb_set_network_header(skb, skb_inner_network_offset(skb));
52 skb_set_transport_header(skb, skb_inner_transport_offset(skb));
53 skb->mac_len = skb_inner_network_offset(skb);
54 skb->protocol = new_protocol;
55
56 need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
57 skb->encap_hdr_csum = need_csum;
58
59 remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
60 skb->remcsum_offload = remcsum;
61
62 need_ipsec = skb_dst(skb) && dst_xfrm(skb_dst(skb));
63 /* Try to offload checksum if possible */
64 offload_csum = !!(need_csum &&
65 !need_ipsec &&
66 (skb->dev->features &
67 (is_ipv6 ? (NETIF_F_HW_CSUM | NETIF_F_IPV6_CSUM) :
68 (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM))));
69
70 features &= skb->dev->hw_enc_features;
71 /* CRC checksum can't be handled by HW when it's a UDP tunneling packet. */
72 features &= ~NETIF_F_SCTP_CRC;
73
74 /* The only checksum offload we care about from here on out is the
75 * outer one so strip the existing checksum feature flags and
76 * instead set the flag based on our outer checksum offload value.
77 */
78 if (remcsum) {
79 features &= ~NETIF_F_CSUM_MASK;
80 if (!need_csum || offload_csum)
81 features |= NETIF_F_HW_CSUM;
82 }
83
84 /* segment inner packet. */
85 segs = gso_inner_segment(skb, features);
86 if (IS_ERR_OR_NULL(segs)) {
87 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
88 mac_len);
89 goto out;
90 }
91
92 gso_partial = !!(skb_shinfo(segs)->gso_type & SKB_GSO_PARTIAL);
93
94 outer_hlen = skb_tnl_header_len(skb);
95 udp_offset = outer_hlen - tnl_hlen;
96 skb = segs;
97 do {
98 unsigned int len;
99
100 if (remcsum)
101 skb->ip_summed = CHECKSUM_NONE;
102
103 /* Set up inner headers if we are offloading inner checksum */
104 if (skb->ip_summed == CHECKSUM_PARTIAL) {
105 skb_reset_inner_headers(skb);
106 skb->encapsulation = 1;
107 }
108
109 skb->mac_len = mac_len;
110 skb->protocol = protocol;
111
112 __skb_push(skb, outer_hlen);
113 skb_reset_mac_header(skb);
114 skb_set_network_header(skb, mac_len);
115 skb_set_transport_header(skb, udp_offset);
116 len = skb->len - udp_offset;
117 uh = udp_hdr(skb);
118
119 /* If we are only performing partial GSO the inner header
120 * will be using a length value equal to only one MSS sized
121 * segment instead of the entire frame.
122 */
123 if (gso_partial && skb_is_gso(skb)) {
124 uh->len = htons(skb_shinfo(skb)->gso_size +
125 SKB_GSO_CB(skb)->data_offset +
126 skb->head - (unsigned char *)uh);
127 } else {
128 uh->len = htons(len);
129 }
130
131 if (!need_csum)
132 continue;
133
134 uh->check = ~csum_fold(csum_add(partial,
135 (__force __wsum)htonl(len)));
136
137 if (skb->encapsulation || !offload_csum) {
138 uh->check = gso_make_checksum(skb, ~uh->check);
139 if (uh->check == 0)
140 uh->check = CSUM_MANGLED_0;
141 } else {
142 skb->ip_summed = CHECKSUM_PARTIAL;
143 skb->csum_start = skb_transport_header(skb) - skb->head;
144 skb->csum_offset = offsetof(struct udphdr, check);
145 }
146 } while ((skb = skb->next));
147 out:
148 return segs;
149 }
150
151 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
152 netdev_features_t features,
153 bool is_ipv6)
154 {
155 __be16 protocol = skb->protocol;
156 const struct net_offload **offloads;
157 const struct net_offload *ops;
158 struct sk_buff *segs = ERR_PTR(-EINVAL);
159 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
160 netdev_features_t features);
161
162 rcu_read_lock();
163
164 switch (skb->inner_protocol_type) {
165 case ENCAP_TYPE_ETHER:
166 protocol = skb->inner_protocol;
167 gso_inner_segment = skb_mac_gso_segment;
168 break;
169 case ENCAP_TYPE_IPPROTO:
170 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
171 ops = rcu_dereference(offloads[skb->inner_ipproto]);
172 if (!ops || !ops->callbacks.gso_segment)
173 goto out_unlock;
174 gso_inner_segment = ops->callbacks.gso_segment;
175 break;
176 default:
177 goto out_unlock;
178 }
179
180 segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
181 protocol, is_ipv6);
182
183 out_unlock:
184 rcu_read_unlock();
185
186 return segs;
187 }
188 EXPORT_SYMBOL(skb_udp_tunnel_segment);
189
190 static struct sk_buff *__udp_gso_segment_list(struct sk_buff *skb,
191 netdev_features_t features)
192 {
193 unsigned int mss = skb_shinfo(skb)->gso_size;
194
195 skb = skb_segment_list(skb, features, skb_mac_header_len(skb));
196 if (IS_ERR(skb))
197 return skb;
198
199 udp_hdr(skb)->len = htons(sizeof(struct udphdr) + mss);
200
201 return skb;
202 }
203
204 struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
205 netdev_features_t features)
206 {
207 struct sock *sk = gso_skb->sk;
208 unsigned int sum_truesize = 0;
209 struct sk_buff *segs, *seg;
210 struct udphdr *uh;
211 unsigned int mss;
212 bool copy_dtor;
213 __sum16 check;
214 __be16 newlen;
215
216 if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST)
217 return __udp_gso_segment_list(gso_skb, features);
218
219 mss = skb_shinfo(gso_skb)->gso_size;
220 if (gso_skb->len <= sizeof(*uh) + mss)
221 return ERR_PTR(-EINVAL);
222
223 skb_pull(gso_skb, sizeof(*uh));
224
225 /* clear destructor to avoid skb_segment assigning it to tail */
226 copy_dtor = gso_skb->destructor == sock_wfree;
227 if (copy_dtor)
228 gso_skb->destructor = NULL;
229
230 segs = skb_segment(gso_skb, features);
231 if (IS_ERR_OR_NULL(segs)) {
232 if (copy_dtor)
233 gso_skb->destructor = sock_wfree;
234 return segs;
235 }
236
237 /* GSO partial and frag_list segmentation only requires splitting
238 * the frame into an MSS multiple and possibly a remainder, both
239 * cases return a GSO skb. So update the mss now.
240 */
241 if (skb_is_gso(segs))
242 mss *= skb_shinfo(segs)->gso_segs;
243
244 seg = segs;
245 uh = udp_hdr(seg);
246
247 /* preserve TX timestamp flags and TS key for first segment */
248 skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey;
249 skb_shinfo(seg)->tx_flags |=
250 (skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP);
251
252 /* compute checksum adjustment based on old length versus new */
253 newlen = htons(sizeof(*uh) + mss);
254 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
255
256 for (;;) {
257 if (copy_dtor) {
258 seg->destructor = sock_wfree;
259 seg->sk = sk;
260 sum_truesize += seg->truesize;
261 }
262
263 if (!seg->next)
264 break;
265
266 uh->len = newlen;
267 uh->check = check;
268
269 if (seg->ip_summed == CHECKSUM_PARTIAL)
270 gso_reset_checksum(seg, ~check);
271 else
272 uh->check = gso_make_checksum(seg, ~check) ? :
273 CSUM_MANGLED_0;
274
275 seg = seg->next;
276 uh = udp_hdr(seg);
277 }
278
279 /* last packet can be partial gso_size, account for that in checksum */
280 newlen = htons(skb_tail_pointer(seg) - skb_transport_header(seg) +
281 seg->data_len);
282 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
283
284 uh->len = newlen;
285 uh->check = check;
286
287 if (seg->ip_summed == CHECKSUM_PARTIAL)
288 gso_reset_checksum(seg, ~check);
289 else
290 uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0;
291
292 /* update refcount for the packet */
293 if (copy_dtor) {
294 int delta = sum_truesize - gso_skb->truesize;
295
296 /* In some pathological cases, delta can be negative.
297 * We need to either use refcount_add() or refcount_sub_and_test()
298 */
299 if (likely(delta >= 0))
300 refcount_add(delta, &sk->sk_wmem_alloc);
301 else
302 WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));
303 }
304 return segs;
305 }
306 EXPORT_SYMBOL_GPL(__udp_gso_segment);
307
308 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
309 netdev_features_t features)
310 {
311 struct sk_buff *segs = ERR_PTR(-EINVAL);
312 unsigned int mss;
313 __wsum csum;
314 struct udphdr *uh;
315 struct iphdr *iph;
316
317 if (skb->encapsulation &&
318 (skb_shinfo(skb)->gso_type &
319 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
320 segs = skb_udp_tunnel_segment(skb, features, false);
321 goto out;
322 }
323
324 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4)))
325 goto out;
326
327 if (!pskb_may_pull(skb, sizeof(struct udphdr)))
328 goto out;
329
330 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
331 return __udp_gso_segment(skb, features);
332
333 mss = skb_shinfo(skb)->gso_size;
334 if (unlikely(skb->len <= mss))
335 goto out;
336
337 /* Do software UFO. Complete and fill in the UDP checksum as
338 * HW cannot do checksum of UDP packets sent as multiple
339 * IP fragments.
340 */
341
342 uh = udp_hdr(skb);
343 iph = ip_hdr(skb);
344
345 uh->check = 0;
346 csum = skb_checksum(skb, 0, skb->len, 0);
347 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
348 if (uh->check == 0)
349 uh->check = CSUM_MANGLED_0;
350
351 skb->ip_summed = CHECKSUM_UNNECESSARY;
352
353 /* If there is no outer header we can fake a checksum offload
354 * due to the fact that we have already done the checksum in
355 * software prior to segmenting the frame.
356 */
357 if (!skb->encap_hdr_csum)
358 features |= NETIF_F_HW_CSUM;
359
360 /* Fragment the skb. IP headers of the fragments are updated in
361 * inet_gso_segment()
362 */
363 segs = skb_segment(skb, features);
364 out:
365 return segs;
366 }
367
368 #define UDP_GRO_CNT_MAX 64
369 static struct sk_buff *udp_gro_receive_segment(struct list_head *head,
370 struct sk_buff *skb)
371 {
372 struct udphdr *uh = udp_gro_udphdr(skb);
373 struct sk_buff *pp = NULL;
374 struct udphdr *uh2;
375 struct sk_buff *p;
376 unsigned int ulen;
377 int ret = 0;
378
379 /* requires non zero csum, for symmetry with GSO */
380 if (!uh->check) {
381 NAPI_GRO_CB(skb)->flush = 1;
382 return NULL;
383 }
384
385 /* Do not deal with padded or malicious packets, sorry ! */
386 ulen = ntohs(uh->len);
387 if (ulen <= sizeof(*uh) || ulen != skb_gro_len(skb)) {
388 NAPI_GRO_CB(skb)->flush = 1;
389 return NULL;
390 }
391 /* pull encapsulating udp header */
392 skb_gro_pull(skb, sizeof(struct udphdr));
393
394 list_for_each_entry(p, head, list) {
395 if (!NAPI_GRO_CB(p)->same_flow)
396 continue;
397
398 uh2 = udp_hdr(p);
399
400 /* Match ports only, as csum is always non zero */
401 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) {
402 NAPI_GRO_CB(p)->same_flow = 0;
403 continue;
404 }
405
406 if (NAPI_GRO_CB(skb)->is_flist != NAPI_GRO_CB(p)->is_flist) {
407 NAPI_GRO_CB(skb)->flush = 1;
408 return p;
409 }
410
411 /* Terminate the flow on len mismatch or if it grow "too much".
412 * Under small packet flood GRO count could elsewhere grow a lot
413 * leading to excessive truesize values.
414 * On len mismatch merge the first packet shorter than gso_size,
415 * otherwise complete the GRO packet.
416 */
417 if (ulen > ntohs(uh2->len)) {
418 pp = p;
419 } else {
420 if (NAPI_GRO_CB(skb)->is_flist) {
421 if (!pskb_may_pull(skb, skb_gro_offset(skb))) {
422 NAPI_GRO_CB(skb)->flush = 1;
423 return NULL;
424 }
425 if ((skb->ip_summed != p->ip_summed) ||
426 (skb->csum_level != p->csum_level)) {
427 NAPI_GRO_CB(skb)->flush = 1;
428 return NULL;
429 }
430 ret = skb_gro_receive_list(p, skb);
431 } else {
432 skb_gro_postpull_rcsum(skb, uh,
433 sizeof(struct udphdr));
434
435 ret = skb_gro_receive(p, skb);
436 }
437 }
438
439 if (ret || ulen != ntohs(uh2->len) ||
440 NAPI_GRO_CB(p)->count >= UDP_GRO_CNT_MAX)
441 pp = p;
442
443 return pp;
444 }
445
446 /* mismatch, but we never need to flush */
447 return NULL;
448 }
449
450 struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb,
451 struct udphdr *uh, struct sock *sk)
452 {
453 struct sk_buff *pp = NULL;
454 struct sk_buff *p;
455 struct udphdr *uh2;
456 unsigned int off = skb_gro_offset(skb);
457 int flush = 1;
458
459 NAPI_GRO_CB(skb)->is_flist = 0;
460 if (skb->dev->features & NETIF_F_GRO_FRAGLIST)
461 NAPI_GRO_CB(skb)->is_flist = sk ? !udp_sk(sk)->gro_enabled: 1;
462
463 if ((sk && udp_sk(sk)->gro_enabled) || NAPI_GRO_CB(skb)->is_flist) {
464 pp = call_gro_receive(udp_gro_receive_segment, head, skb);
465 return pp;
466 }
467
468 if (!sk || NAPI_GRO_CB(skb)->encap_mark ||
469 (skb->ip_summed != CHECKSUM_PARTIAL &&
470 NAPI_GRO_CB(skb)->csum_cnt == 0 &&
471 !NAPI_GRO_CB(skb)->csum_valid) ||
472 !udp_sk(sk)->gro_receive)
473 goto out;
474
475 /* mark that this skb passed once through the tunnel gro layer */
476 NAPI_GRO_CB(skb)->encap_mark = 1;
477
478 flush = 0;
479
480 list_for_each_entry(p, head, list) {
481 if (!NAPI_GRO_CB(p)->same_flow)
482 continue;
483
484 uh2 = (struct udphdr *)(p->data + off);
485
486 /* Match ports and either checksums are either both zero
487 * or nonzero.
488 */
489 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
490 (!uh->check ^ !uh2->check)) {
491 NAPI_GRO_CB(p)->same_flow = 0;
492 continue;
493 }
494 }
495
496 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
497 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
498 pp = call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
499
500 out:
501 skb_gro_flush_final(skb, pp, flush);
502 return pp;
503 }
504 EXPORT_SYMBOL(udp_gro_receive);
505
506 static struct sock *udp4_gro_lookup_skb(struct sk_buff *skb, __be16 sport,
507 __be16 dport)
508 {
509 const struct iphdr *iph = skb_gro_network_header(skb);
510
511 return __udp4_lib_lookup(dev_net(skb->dev), iph->saddr, sport,
512 iph->daddr, dport, inet_iif(skb),
513 inet_sdif(skb), &udp_table, NULL);
514 }
515
516 INDIRECT_CALLABLE_SCOPE
517 struct sk_buff *udp4_gro_receive(struct list_head *head, struct sk_buff *skb)
518 {
519 struct udphdr *uh = udp_gro_udphdr(skb);
520 struct sock *sk = NULL;
521 struct sk_buff *pp;
522
523 if (unlikely(!uh))
524 goto flush;
525
526 /* Don't bother verifying checksum if we're going to flush anyway. */
527 if (NAPI_GRO_CB(skb)->flush)
528 goto skip;
529
530 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
531 inet_gro_compute_pseudo))
532 goto flush;
533 else if (uh->check)
534 skb_gro_checksum_try_convert(skb, IPPROTO_UDP,
535 inet_gro_compute_pseudo);
536 skip:
537 NAPI_GRO_CB(skb)->is_ipv6 = 0;
538 rcu_read_lock();
539
540 if (static_branch_unlikely(&udp_encap_needed_key))
541 sk = udp4_gro_lookup_skb(skb, uh->source, uh->dest);
542
543 pp = udp_gro_receive(head, skb, uh, sk);
544 rcu_read_unlock();
545 return pp;
546
547 flush:
548 NAPI_GRO_CB(skb)->flush = 1;
549 return NULL;
550 }
551
552 static int udp_gro_complete_segment(struct sk_buff *skb)
553 {
554 struct udphdr *uh = udp_hdr(skb);
555
556 skb->csum_start = (unsigned char *)uh - skb->head;
557 skb->csum_offset = offsetof(struct udphdr, check);
558 skb->ip_summed = CHECKSUM_PARTIAL;
559
560 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
561 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_L4;
562 return 0;
563 }
564
565 int udp_gro_complete(struct sk_buff *skb, int nhoff,
566 udp_lookup_t lookup)
567 {
568 __be16 newlen = htons(skb->len - nhoff);
569 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
570 struct sock *sk;
571 int err;
572
573 uh->len = newlen;
574
575 rcu_read_lock();
576 sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb,
577 udp4_lib_lookup_skb, skb, uh->source, uh->dest);
578 if (sk && udp_sk(sk)->gro_complete) {
579 skb_shinfo(skb)->gso_type = uh->check ? SKB_GSO_UDP_TUNNEL_CSUM
580 : SKB_GSO_UDP_TUNNEL;
581
582 /* Set encapsulation before calling into inner gro_complete()
583 * functions to make them set up the inner offsets.
584 */
585 skb->encapsulation = 1;
586 err = udp_sk(sk)->gro_complete(sk, skb,
587 nhoff + sizeof(struct udphdr));
588 } else {
589 err = udp_gro_complete_segment(skb);
590 }
591 rcu_read_unlock();
592
593 if (skb->remcsum_offload)
594 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
595
596 return err;
597 }
598 EXPORT_SYMBOL(udp_gro_complete);
599
600 INDIRECT_CALLABLE_SCOPE int udp4_gro_complete(struct sk_buff *skb, int nhoff)
601 {
602 const struct iphdr *iph = ip_hdr(skb);
603 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
604
605 if (NAPI_GRO_CB(skb)->is_flist) {
606 uh->len = htons(skb->len - nhoff);
607
608 skb_shinfo(skb)->gso_type |= (SKB_GSO_FRAGLIST|SKB_GSO_UDP_L4);
609 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
610
611 if (skb->ip_summed == CHECKSUM_UNNECESSARY) {
612 if (skb->csum_level < SKB_MAX_CSUM_LEVEL)
613 skb->csum_level++;
614 } else {
615 skb->ip_summed = CHECKSUM_UNNECESSARY;
616 skb->csum_level = 0;
617 }
618
619 return 0;
620 }
621
622 if (uh->check)
623 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
624 iph->daddr, 0);
625
626 return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
627 }
628
629 static const struct net_offload udpv4_offload = {
630 .callbacks = {
631 .gso_segment = udp4_ufo_fragment,
632 .gro_receive = udp4_gro_receive,
633 .gro_complete = udp4_gro_complete,
634 },
635 };
636
637 int __init udpv4_offload_init(void)
638 {
639 return inet_add_offload(&udpv4_offload, IPPROTO_UDP);
640 }
Linux with added FPC-III support