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