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 static struct sk_buff *__udp_gso_segment_list(struct sk_buff *skb, 188 netdev_features_t features) 189 { 190 unsigned int mss = skb_shinfo(skb)->gso_size; 191 192 skb = skb_segment_list(skb, features, skb_mac_header_len(skb)); 193 if (IS_ERR(skb)) 194 return skb; 195 196 udp_hdr(skb)->len = htons(sizeof(struct udphdr) + mss); 197 198 return skb; 199 } 200 201 struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb, 202 netdev_features_t features) 203 { 204 struct sock *sk = gso_skb->sk; 205 unsigned int sum_truesize = 0; 206 struct sk_buff *segs, *seg; 207 struct udphdr *uh; 208 unsigned int mss; 209 bool copy_dtor; 210 __sum16 check; 211 __be16 newlen; 212 213 if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST) 214 return __udp_gso_segment_list(gso_skb, features); 215 216 mss = skb_shinfo(gso_skb)->gso_size; 217 if (gso_skb->len <= sizeof(*uh) + mss) 218 return ERR_PTR(-EINVAL); 219 220 skb_pull(gso_skb, sizeof(*uh)); 221 222 /* clear destructor to avoid skb_segment assigning it to tail */ 223 copy_dtor = gso_skb->destructor == sock_wfree; 224 if (copy_dtor) 225 gso_skb->destructor = NULL; 226 227 segs = skb_segment(gso_skb, features); 228 if (IS_ERR_OR_NULL(segs)) { 229 if (copy_dtor) 230 gso_skb->destructor = sock_wfree; 231 return segs; 232 } 233 234 /* GSO partial and frag_list segmentation only requires splitting 235 * the frame into an MSS multiple and possibly a remainder, both 236 * cases return a GSO skb. So update the mss now. 237 */ 238 if (skb_is_gso(segs)) 239 mss *= skb_shinfo(segs)->gso_segs; 240 241 seg = segs; 242 uh = udp_hdr(seg); 243 244 /* preserve TX timestamp flags and TS key for first segment */ 245 skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey; 246 skb_shinfo(seg)->tx_flags |= 247 (skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP); 248 249 /* compute checksum adjustment based on old length versus new */ 250 newlen = htons(sizeof(*uh) + mss); 251 check = csum16_add(csum16_sub(uh->check, uh->len), newlen); 252 253 for (;;) { 254 if (copy_dtor) { 255 seg->destructor = sock_wfree; 256 seg->sk = sk; 257 sum_truesize += seg->truesize; 258 } 259 260 if (!seg->next) 261 break; 262 263 uh->len = newlen; 264 uh->check = check; 265 266 if (seg->ip_summed == CHECKSUM_PARTIAL) 267 gso_reset_checksum(seg, ~check); 268 else 269 uh->check = gso_make_checksum(seg, ~check) ? : 270 CSUM_MANGLED_0; 271 272 seg = seg->next; 273 uh = udp_hdr(seg); 274 } 275 276 /* last packet can be partial gso_size, account for that in checksum */ 277 newlen = htons(skb_tail_pointer(seg) - skb_transport_header(seg) + 278 seg->data_len); 279 check = csum16_add(csum16_sub(uh->check, uh->len), newlen); 280 281 uh->len = newlen; 282 uh->check = check; 283 284 if (seg->ip_summed == CHECKSUM_PARTIAL) 285 gso_reset_checksum(seg, ~check); 286 else 287 uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0; 288 289 /* update refcount for the packet */ 290 if (copy_dtor) { 291 int delta = sum_truesize - gso_skb->truesize; 292 293 /* In some pathological cases, delta can be negative. 294 * We need to either use refcount_add() or refcount_sub_and_test() 295 */ 296 if (likely(delta >= 0)) 297 refcount_add(delta, &sk->sk_wmem_alloc); 298 else 299 WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc)); 300 } 301 return segs; 302 } 303 EXPORT_SYMBOL_GPL(__udp_gso_segment); 304 305 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb, 306 netdev_features_t features) 307 { 308 struct sk_buff *segs = ERR_PTR(-EINVAL); 309 unsigned int mss; 310 __wsum csum; 311 struct udphdr *uh; 312 struct iphdr *iph; 313 314 if (skb->encapsulation && 315 (skb_shinfo(skb)->gso_type & 316 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) { 317 segs = skb_udp_tunnel_segment(skb, features, false); 318 goto out; 319 } 320 321 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4))) 322 goto out; 323 324 if (!pskb_may_pull(skb, sizeof(struct udphdr))) 325 goto out; 326 327 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) 328 return __udp_gso_segment(skb, features); 329 330 mss = skb_shinfo(skb)->gso_size; 331 if (unlikely(skb->len <= mss)) 332 goto out; 333 334 /* Do software UFO. Complete and fill in the UDP checksum as 335 * HW cannot do checksum of UDP packets sent as multiple 336 * IP fragments. 337 */ 338 339 uh = udp_hdr(skb); 340 iph = ip_hdr(skb); 341 342 uh->check = 0; 343 csum = skb_checksum(skb, 0, skb->len, 0); 344 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum); 345 if (uh->check == 0) 346 uh->check = CSUM_MANGLED_0; 347 348 skb->ip_summed = CHECKSUM_UNNECESSARY; 349 350 /* If there is no outer header we can fake a checksum offload 351 * due to the fact that we have already done the checksum in 352 * software prior to segmenting the frame. 353 */ 354 if (!skb->encap_hdr_csum) 355 features |= NETIF_F_HW_CSUM; 356 357 /* Fragment the skb. IP headers of the fragments are updated in 358 * inet_gso_segment() 359 */ 360 segs = skb_segment(skb, features); 361 out: 362 return segs; 363 } 364 365 #define UDP_GRO_CNT_MAX 64 366 static struct sk_buff *udp_gro_receive_segment(struct list_head *head, 367 struct sk_buff *skb) 368 { 369 struct udphdr *uh = udp_hdr(skb); 370 struct sk_buff *pp = NULL; 371 struct udphdr *uh2; 372 struct sk_buff *p; 373 unsigned int ulen; 374 int ret = 0; 375 376 /* requires non zero csum, for symmetry with GSO */ 377 if (!uh->check) { 378 NAPI_GRO_CB(skb)->flush = 1; 379 return NULL; 380 } 381 382 /* Do not deal with padded or malicious packets, sorry ! */ 383 ulen = ntohs(uh->len); 384 if (ulen <= sizeof(*uh) || ulen != skb_gro_len(skb)) { 385 NAPI_GRO_CB(skb)->flush = 1; 386 return NULL; 387 } 388 /* pull encapsulating udp header */ 389 skb_gro_pull(skb, sizeof(struct udphdr)); 390 391 list_for_each_entry(p, head, list) { 392 if (!NAPI_GRO_CB(p)->same_flow) 393 continue; 394 395 uh2 = udp_hdr(p); 396 397 /* Match ports only, as csum is always non zero */ 398 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) { 399 NAPI_GRO_CB(p)->same_flow = 0; 400 continue; 401 } 402 403 if (NAPI_GRO_CB(skb)->is_flist != NAPI_GRO_CB(p)->is_flist) { 404 NAPI_GRO_CB(skb)->flush = 1; 405 return p; 406 } 407 408 /* Terminate the flow on len mismatch or if it grow "too much". 409 * Under small packet flood GRO count could elsewhere grow a lot 410 * leading to excessive truesize values. 411 * On len mismatch merge the first packet shorter than gso_size, 412 * otherwise complete the GRO packet. 413 */ 414 if (ulen > ntohs(uh2->len)) { 415 pp = p; 416 } else { 417 if (NAPI_GRO_CB(skb)->is_flist) { 418 if (!pskb_may_pull(skb, skb_gro_offset(skb))) { 419 NAPI_GRO_CB(skb)->flush = 1; 420 return NULL; 421 } 422 if ((skb->ip_summed != p->ip_summed) || 423 (skb->csum_level != p->csum_level)) { 424 NAPI_GRO_CB(skb)->flush = 1; 425 return NULL; 426 } 427 ret = skb_gro_receive_list(p, skb); 428 } else { 429 skb_gro_postpull_rcsum(skb, uh, 430 sizeof(struct udphdr)); 431 432 ret = skb_gro_receive(p, skb); 433 } 434 } 435 436 if (ret || ulen != ntohs(uh2->len) || 437 NAPI_GRO_CB(p)->count >= UDP_GRO_CNT_MAX) 438 pp = p; 439 440 return pp; 441 } 442 443 /* mismatch, but we never need to flush */ 444 return NULL; 445 } 446 447 struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb, 448 struct udphdr *uh, struct sock *sk) 449 { 450 struct sk_buff *pp = NULL; 451 struct sk_buff *p; 452 struct udphdr *uh2; 453 unsigned int off = skb_gro_offset(skb); 454 int flush = 1; 455 456 NAPI_GRO_CB(skb)->is_flist = 0; 457 if (skb->dev->features & NETIF_F_GRO_FRAGLIST) 458 NAPI_GRO_CB(skb)->is_flist = sk ? !udp_sk(sk)->gro_enabled: 1; 459 460 if ((sk && udp_sk(sk)->gro_enabled) || NAPI_GRO_CB(skb)->is_flist) { 461 pp = call_gro_receive(udp_gro_receive_segment, head, skb); 462 return pp; 463 } 464 465 if (!sk || NAPI_GRO_CB(skb)->encap_mark || 466 (skb->ip_summed != CHECKSUM_PARTIAL && 467 NAPI_GRO_CB(skb)->csum_cnt == 0 && 468 !NAPI_GRO_CB(skb)->csum_valid) || 469 !udp_sk(sk)->gro_receive) 470 goto out; 471 472 /* mark that this skb passed once through the tunnel gro layer */ 473 NAPI_GRO_CB(skb)->encap_mark = 1; 474 475 flush = 0; 476 477 list_for_each_entry(p, head, list) { 478 if (!NAPI_GRO_CB(p)->same_flow) 479 continue; 480 481 uh2 = (struct udphdr *)(p->data + off); 482 483 /* Match ports and either checksums are either both zero 484 * or nonzero. 485 */ 486 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) || 487 (!uh->check ^ !uh2->check)) { 488 NAPI_GRO_CB(p)->same_flow = 0; 489 continue; 490 } 491 } 492 493 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */ 494 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr)); 495 pp = call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb); 496 497 out: 498 skb_gro_flush_final(skb, pp, flush); 499 return pp; 500 } 501 EXPORT_SYMBOL(udp_gro_receive); 502 503 INDIRECT_CALLABLE_SCOPE 504 struct sk_buff *udp4_gro_receive(struct list_head *head, struct sk_buff *skb) 505 { 506 struct udphdr *uh = udp_gro_udphdr(skb); 507 struct sk_buff *pp; 508 struct sock *sk; 509 510 if (unlikely(!uh)) 511 goto flush; 512 513 /* Don't bother verifying checksum if we're going to flush anyway. */ 514 if (NAPI_GRO_CB(skb)->flush) 515 goto skip; 516 517 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check, 518 inet_gro_compute_pseudo)) 519 goto flush; 520 else if (uh->check) 521 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, 522 inet_gro_compute_pseudo); 523 skip: 524 NAPI_GRO_CB(skb)->is_ipv6 = 0; 525 rcu_read_lock(); 526 sk = static_branch_unlikely(&udp_encap_needed_key) ? udp4_lib_lookup_skb(skb, uh->source, uh->dest) : NULL; 527 pp = udp_gro_receive(head, skb, uh, sk); 528 rcu_read_unlock(); 529 return pp; 530 531 flush: 532 NAPI_GRO_CB(skb)->flush = 1; 533 return NULL; 534 } 535 536 static int udp_gro_complete_segment(struct sk_buff *skb) 537 { 538 struct udphdr *uh = udp_hdr(skb); 539 540 skb->csum_start = (unsigned char *)uh - skb->head; 541 skb->csum_offset = offsetof(struct udphdr, check); 542 skb->ip_summed = CHECKSUM_PARTIAL; 543 544 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count; 545 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_L4; 546 return 0; 547 } 548 549 int udp_gro_complete(struct sk_buff *skb, int nhoff, 550 udp_lookup_t lookup) 551 { 552 __be16 newlen = htons(skb->len - nhoff); 553 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff); 554 int err = -ENOSYS; 555 struct sock *sk; 556 557 uh->len = newlen; 558 559 rcu_read_lock(); 560 sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb, 561 udp4_lib_lookup_skb, skb, uh->source, uh->dest); 562 if (sk && udp_sk(sk)->gro_complete) { 563 skb_shinfo(skb)->gso_type = uh->check ? SKB_GSO_UDP_TUNNEL_CSUM 564 : SKB_GSO_UDP_TUNNEL; 565 566 /* Set encapsulation before calling into inner gro_complete() 567 * functions to make them set up the inner offsets. 568 */ 569 skb->encapsulation = 1; 570 err = udp_sk(sk)->gro_complete(sk, skb, 571 nhoff + sizeof(struct udphdr)); 572 } else { 573 err = udp_gro_complete_segment(skb); 574 } 575 rcu_read_unlock(); 576 577 if (skb->remcsum_offload) 578 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM; 579 580 return err; 581 } 582 EXPORT_SYMBOL(udp_gro_complete); 583 584 INDIRECT_CALLABLE_SCOPE int udp4_gro_complete(struct sk_buff *skb, int nhoff) 585 { 586 const struct iphdr *iph = ip_hdr(skb); 587 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff); 588 589 if (NAPI_GRO_CB(skb)->is_flist) { 590 uh->len = htons(skb->len - nhoff); 591 592 skb_shinfo(skb)->gso_type |= (SKB_GSO_FRAGLIST|SKB_GSO_UDP_L4); 593 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count; 594 595 if (skb->ip_summed == CHECKSUM_UNNECESSARY) { 596 if (skb->csum_level < SKB_MAX_CSUM_LEVEL) 597 skb->csum_level++; 598 } else { 599 skb->ip_summed = CHECKSUM_UNNECESSARY; 600 skb->csum_level = 0; 601 } 602 603 return 0; 604 } 605 606 if (uh->check) 607 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr, 608 iph->daddr, 0); 609 610 return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb); 611 } 612 613 static const struct net_offload udpv4_offload = { 614 .callbacks = { 615 .gso_segment = udp4_ufo_fragment, 616 .gro_receive = udp4_gro_receive, 617 .gro_complete = udp4_gro_complete, 618 }, 619 }; 620 621 int __init udpv4_offload_init(void) 622 { 623 return inet_add_offload(&udpv4_offload, IPPROTO_UDP); 624 } 625