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/gro.h> 11 #include <net/gso.h> 12 #include <net/udp.h> 13 #include <net/protocol.h> 14 #include <net/inet_common.h> 15 #include <net/udp_tunnel.h> 16 17 #if IS_ENABLED(CONFIG_NET_UDP_TUNNEL) 18 19 /* 20 * Dummy GRO tunnel callback, exists mainly to avoid dangling/NULL 21 * values for the udp tunnel static call. 22 */ 23 static struct sk_buff *dummy_gro_rcv(struct sock *sk, 24 struct list_head *head, 25 struct sk_buff *skb) 26 { 27 NAPI_GRO_CB(skb)->flush = 1; 28 return NULL; 29 } 30 31 typedef struct sk_buff *(*udp_tunnel_gro_rcv_t)(struct sock *sk, 32 struct list_head *head, 33 struct sk_buff *skb); 34 35 struct udp_tunnel_type_entry { 36 udp_tunnel_gro_rcv_t gro_receive; 37 refcount_t count; 38 }; 39 40 #define UDP_MAX_TUNNEL_TYPES (IS_ENABLED(CONFIG_GENEVE) + \ 41 IS_ENABLED(CONFIG_VXLAN) * 2 + \ 42 IS_ENABLED(CONFIG_NET_FOU) * 2 + \ 43 IS_ENABLED(CONFIG_XFRM) * 2) 44 45 DEFINE_STATIC_CALL(udp_tunnel_gro_rcv, dummy_gro_rcv); 46 static DEFINE_STATIC_KEY_FALSE(udp_tunnel_static_call); 47 static DEFINE_MUTEX(udp_tunnel_gro_type_lock); 48 static struct udp_tunnel_type_entry udp_tunnel_gro_types[UDP_MAX_TUNNEL_TYPES]; 49 static unsigned int udp_tunnel_gro_type_nr; 50 static DEFINE_SPINLOCK(udp_tunnel_gro_lock); 51 52 void udp_tunnel_update_gro_lookup(struct net *net, struct sock *sk, bool add) 53 { 54 bool is_ipv6 = sk->sk_family == AF_INET6; 55 struct udp_sock *tup, *up = udp_sk(sk); 56 struct udp_tunnel_gro *udp_tunnel_gro; 57 58 spin_lock(&udp_tunnel_gro_lock); 59 udp_tunnel_gro = &net->ipv4.udp_tunnel_gro[is_ipv6]; 60 if (add) 61 hlist_add_head(&up->tunnel_list, &udp_tunnel_gro->list); 62 else if (up->tunnel_list.pprev) 63 hlist_del_init(&up->tunnel_list); 64 65 if (udp_tunnel_gro->list.first && 66 !udp_tunnel_gro->list.first->next) { 67 tup = hlist_entry(udp_tunnel_gro->list.first, struct udp_sock, 68 tunnel_list); 69 70 rcu_assign_pointer(udp_tunnel_gro->sk, (struct sock *)tup); 71 } else { 72 RCU_INIT_POINTER(udp_tunnel_gro->sk, NULL); 73 } 74 75 spin_unlock(&udp_tunnel_gro_lock); 76 } 77 EXPORT_SYMBOL_GPL(udp_tunnel_update_gro_lookup); 78 79 void udp_tunnel_update_gro_rcv(struct sock *sk, bool add) 80 { 81 struct udp_tunnel_type_entry *cur = NULL; 82 struct udp_sock *up = udp_sk(sk); 83 int i, old_gro_type_nr; 84 85 if (!UDP_MAX_TUNNEL_TYPES || !up->gro_receive) 86 return; 87 88 mutex_lock(&udp_tunnel_gro_type_lock); 89 90 /* Check if the static call is permanently disabled. */ 91 if (udp_tunnel_gro_type_nr > UDP_MAX_TUNNEL_TYPES) 92 goto out; 93 94 for (i = 0; i < udp_tunnel_gro_type_nr; i++) 95 if (udp_tunnel_gro_types[i].gro_receive == up->gro_receive) 96 cur = &udp_tunnel_gro_types[i]; 97 98 old_gro_type_nr = udp_tunnel_gro_type_nr; 99 if (add) { 100 /* 101 * Update the matching entry, if found, or add a new one 102 * if needed 103 */ 104 if (cur) { 105 refcount_inc(&cur->count); 106 goto out; 107 } 108 109 if (unlikely(udp_tunnel_gro_type_nr == UDP_MAX_TUNNEL_TYPES)) { 110 pr_err_once("Too many UDP tunnel types, please increase UDP_MAX_TUNNEL_TYPES\n"); 111 /* Ensure static call will never be enabled */ 112 udp_tunnel_gro_type_nr = UDP_MAX_TUNNEL_TYPES + 1; 113 } else { 114 cur = &udp_tunnel_gro_types[udp_tunnel_gro_type_nr++]; 115 refcount_set(&cur->count, 1); 116 cur->gro_receive = up->gro_receive; 117 } 118 } else { 119 /* 120 * The stack cleanups only successfully added tunnel, the 121 * lookup on removal should never fail. 122 */ 123 if (WARN_ON_ONCE(!cur)) 124 goto out; 125 126 if (!refcount_dec_and_test(&cur->count)) 127 goto out; 128 129 /* Avoid gaps, so that the enable tunnel has always id 0 */ 130 *cur = udp_tunnel_gro_types[--udp_tunnel_gro_type_nr]; 131 } 132 133 if (udp_tunnel_gro_type_nr == 1) { 134 static_call_update(udp_tunnel_gro_rcv, 135 udp_tunnel_gro_types[0].gro_receive); 136 static_branch_enable(&udp_tunnel_static_call); 137 } else if (old_gro_type_nr == 1) { 138 static_branch_disable(&udp_tunnel_static_call); 139 static_call_update(udp_tunnel_gro_rcv, dummy_gro_rcv); 140 } 141 142 out: 143 mutex_unlock(&udp_tunnel_gro_type_lock); 144 } 145 EXPORT_SYMBOL_GPL(udp_tunnel_update_gro_rcv); 146 147 static struct sk_buff *udp_tunnel_gro_rcv(struct sock *sk, 148 struct list_head *head, 149 struct sk_buff *skb) 150 { 151 if (static_branch_likely(&udp_tunnel_static_call)) { 152 if (unlikely(gro_recursion_inc_test(skb))) { 153 NAPI_GRO_CB(skb)->flush |= 1; 154 return NULL; 155 } 156 return static_call(udp_tunnel_gro_rcv)(sk, head, skb); 157 } 158 return call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb); 159 } 160 161 #else 162 163 static struct sk_buff *udp_tunnel_gro_rcv(struct sock *sk, 164 struct list_head *head, 165 struct sk_buff *skb) 166 { 167 return call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb); 168 } 169 170 #endif 171 172 static struct sk_buff *__skb_udp_tunnel_segment(struct sk_buff *skb, 173 netdev_features_t features, 174 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb, 175 netdev_features_t features), 176 __be16 new_protocol, bool is_ipv6) 177 { 178 int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb); 179 bool remcsum, need_csum, offload_csum, gso_partial; 180 struct sk_buff *segs = ERR_PTR(-EINVAL); 181 struct udphdr *uh = udp_hdr(skb); 182 u16 mac_offset = skb->mac_header; 183 __be16 protocol = skb->protocol; 184 u16 mac_len = skb->mac_len; 185 int udp_offset, outer_hlen; 186 __wsum partial; 187 bool need_ipsec; 188 189 if (unlikely(!pskb_may_pull(skb, tnl_hlen))) 190 goto out; 191 192 /* Adjust partial header checksum to negate old length. 193 * We cannot rely on the value contained in uh->len as it is 194 * possible that the actual value exceeds the boundaries of the 195 * 16 bit length field due to the header being added outside of an 196 * IP or IPv6 frame that was already limited to 64K - 1. 197 */ 198 if (skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL) 199 partial = (__force __wsum)uh->len; 200 else 201 partial = (__force __wsum)htonl(skb->len); 202 partial = csum_sub(csum_unfold(uh->check), partial); 203 204 /* setup inner skb. */ 205 skb->encapsulation = 0; 206 SKB_GSO_CB(skb)->encap_level = 0; 207 __skb_pull(skb, tnl_hlen); 208 skb_reset_mac_header(skb); 209 skb_set_network_header(skb, skb_inner_network_offset(skb)); 210 skb_set_transport_header(skb, skb_inner_transport_offset(skb)); 211 skb->mac_len = skb_inner_network_offset(skb); 212 skb->protocol = new_protocol; 213 214 need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM); 215 skb->encap_hdr_csum = need_csum; 216 217 remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM); 218 skb->remcsum_offload = remcsum; 219 220 need_ipsec = (skb_dst(skb) && dst_xfrm(skb_dst(skb))) || skb_sec_path(skb); 221 /* Try to offload checksum if possible */ 222 offload_csum = !!(need_csum && 223 !need_ipsec && 224 (skb->dev->features & 225 (is_ipv6 ? (NETIF_F_HW_CSUM | NETIF_F_IPV6_CSUM) : 226 (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM)))); 227 228 features &= skb->dev->hw_enc_features; 229 if (need_csum) 230 features &= ~NETIF_F_SCTP_CRC; 231 232 /* The only checksum offload we care about from here on out is the 233 * outer one so strip the existing checksum feature flags and 234 * instead set the flag based on our outer checksum offload value. 235 */ 236 if (remcsum) { 237 features &= ~NETIF_F_CSUM_MASK; 238 if (!need_csum || offload_csum) 239 features |= NETIF_F_HW_CSUM; 240 } 241 242 /* segment inner packet. */ 243 segs = gso_inner_segment(skb, features); 244 if (IS_ERR_OR_NULL(segs)) { 245 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset, 246 mac_len); 247 goto out; 248 } 249 250 gso_partial = !!(skb_shinfo(segs)->gso_type & SKB_GSO_PARTIAL); 251 252 outer_hlen = skb_tnl_header_len(skb); 253 udp_offset = outer_hlen - tnl_hlen; 254 skb = segs; 255 do { 256 unsigned int len; 257 258 if (remcsum) 259 skb->ip_summed = CHECKSUM_NONE; 260 261 /* Set up inner headers if we are offloading inner checksum */ 262 if (skb->ip_summed == CHECKSUM_PARTIAL) { 263 skb_reset_inner_headers(skb); 264 skb->encapsulation = 1; 265 } 266 267 skb->mac_len = mac_len; 268 skb->protocol = protocol; 269 270 __skb_push(skb, outer_hlen); 271 skb_reset_mac_header(skb); 272 skb_set_network_header(skb, mac_len); 273 skb_set_transport_header(skb, udp_offset); 274 len = skb->len - udp_offset; 275 uh = udp_hdr(skb); 276 277 /* If we are only performing partial GSO the inner header 278 * will be using a length value equal to only one MSS sized 279 * segment instead of the entire frame. 280 */ 281 if (gso_partial && skb_is_gso(skb)) { 282 uh->len = htons(skb_shinfo(skb)->gso_size + 283 SKB_GSO_CB(skb)->data_offset + 284 skb->head - (unsigned char *)uh); 285 } else { 286 uh->len = htons(len); 287 } 288 289 if (!need_csum) 290 continue; 291 292 uh->check = ~csum_fold(csum_add(partial, 293 (__force __wsum)htonl(len))); 294 295 if (skb->encapsulation || !offload_csum) { 296 uh->check = gso_make_checksum(skb, ~uh->check); 297 if (uh->check == 0) 298 uh->check = CSUM_MANGLED_0; 299 } else { 300 skb->ip_summed = CHECKSUM_PARTIAL; 301 skb->csum_start = skb_transport_header(skb) - skb->head; 302 skb->csum_offset = offsetof(struct udphdr, check); 303 } 304 } while ((skb = skb->next)); 305 out: 306 return segs; 307 } 308 309 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb, 310 netdev_features_t features, 311 bool is_ipv6) 312 { 313 const struct net_offload __rcu **offloads; 314 __be16 protocol = skb->protocol; 315 const struct net_offload *ops; 316 struct sk_buff *segs = ERR_PTR(-EINVAL); 317 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb, 318 netdev_features_t features); 319 320 rcu_read_lock(); 321 322 switch (skb->inner_protocol_type) { 323 case ENCAP_TYPE_ETHER: 324 protocol = skb->inner_protocol; 325 gso_inner_segment = skb_mac_gso_segment; 326 break; 327 case ENCAP_TYPE_IPPROTO: 328 offloads = is_ipv6 ? inet6_offloads : inet_offloads; 329 ops = rcu_dereference(offloads[skb->inner_ipproto]); 330 if (!ops || !ops->callbacks.gso_segment) 331 goto out_unlock; 332 gso_inner_segment = ops->callbacks.gso_segment; 333 break; 334 default: 335 goto out_unlock; 336 } 337 338 segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment, 339 protocol, is_ipv6); 340 341 out_unlock: 342 rcu_read_unlock(); 343 344 return segs; 345 } 346 EXPORT_SYMBOL(skb_udp_tunnel_segment); 347 348 static void __udpv4_gso_segment_csum(struct sk_buff *seg, 349 __be32 *oldip, __be32 *newip, 350 __be16 *oldport, __be16 *newport) 351 { 352 struct udphdr *uh; 353 struct iphdr *iph; 354 355 if (*oldip == *newip && *oldport == *newport) 356 return; 357 358 uh = udp_hdr(seg); 359 iph = ip_hdr(seg); 360 361 if (uh->check) { 362 inet_proto_csum_replace4(&uh->check, seg, *oldip, *newip, 363 true); 364 inet_proto_csum_replace2(&uh->check, seg, *oldport, *newport, 365 false); 366 if (!uh->check) 367 uh->check = CSUM_MANGLED_0; 368 } 369 *oldport = *newport; 370 371 csum_replace4(&iph->check, *oldip, *newip); 372 *oldip = *newip; 373 } 374 375 static struct sk_buff *__udpv4_gso_segment_list_csum(struct sk_buff *segs) 376 { 377 struct sk_buff *seg; 378 struct udphdr *uh, *uh2; 379 struct iphdr *iph, *iph2; 380 381 seg = segs; 382 uh = udp_hdr(seg); 383 iph = ip_hdr(seg); 384 385 if ((udp_hdr(seg)->dest == udp_hdr(seg->next)->dest) && 386 (udp_hdr(seg)->source == udp_hdr(seg->next)->source) && 387 (ip_hdr(seg)->daddr == ip_hdr(seg->next)->daddr) && 388 (ip_hdr(seg)->saddr == ip_hdr(seg->next)->saddr)) 389 return segs; 390 391 while ((seg = seg->next)) { 392 uh2 = udp_hdr(seg); 393 iph2 = ip_hdr(seg); 394 395 __udpv4_gso_segment_csum(seg, 396 &iph2->saddr, &iph->saddr, 397 &uh2->source, &uh->source); 398 __udpv4_gso_segment_csum(seg, 399 &iph2->daddr, &iph->daddr, 400 &uh2->dest, &uh->dest); 401 } 402 403 return segs; 404 } 405 406 static void __udpv6_gso_segment_csum(struct sk_buff *seg, 407 struct in6_addr *oldip, 408 const struct in6_addr *newip, 409 __be16 *oldport, __be16 newport) 410 { 411 struct udphdr *uh = udp_hdr(seg); 412 413 if (ipv6_addr_equal(oldip, newip) && *oldport == newport) 414 return; 415 416 if (uh->check) { 417 inet_proto_csum_replace16(&uh->check, seg, oldip->s6_addr32, 418 newip->s6_addr32, true); 419 420 inet_proto_csum_replace2(&uh->check, seg, *oldport, newport, 421 false); 422 if (!uh->check) 423 uh->check = CSUM_MANGLED_0; 424 } 425 426 *oldip = *newip; 427 *oldport = newport; 428 } 429 430 static struct sk_buff *__udpv6_gso_segment_list_csum(struct sk_buff *segs) 431 { 432 const struct ipv6hdr *iph; 433 const struct udphdr *uh; 434 struct ipv6hdr *iph2; 435 struct sk_buff *seg; 436 struct udphdr *uh2; 437 438 seg = segs; 439 uh = udp_hdr(seg); 440 iph = ipv6_hdr(seg); 441 uh2 = udp_hdr(seg->next); 442 iph2 = ipv6_hdr(seg->next); 443 444 if (!(*(const u32 *)&uh->source ^ *(const u32 *)&uh2->source) && 445 ipv6_addr_equal(&iph->saddr, &iph2->saddr) && 446 ipv6_addr_equal(&iph->daddr, &iph2->daddr)) 447 return segs; 448 449 while ((seg = seg->next)) { 450 uh2 = udp_hdr(seg); 451 iph2 = ipv6_hdr(seg); 452 453 __udpv6_gso_segment_csum(seg, &iph2->saddr, &iph->saddr, 454 &uh2->source, uh->source); 455 __udpv6_gso_segment_csum(seg, &iph2->daddr, &iph->daddr, 456 &uh2->dest, uh->dest); 457 } 458 459 return segs; 460 } 461 462 static struct sk_buff *__udp_gso_segment_list(struct sk_buff *skb, 463 netdev_features_t features, 464 bool is_ipv6) 465 { 466 unsigned int mss = skb_shinfo(skb)->gso_size; 467 468 skb = skb_segment_list(skb, features, skb_mac_header_len(skb)); 469 if (IS_ERR(skb)) 470 return skb; 471 472 udp_hdr(skb)->len = htons(sizeof(struct udphdr) + mss); 473 474 if (is_ipv6) 475 return __udpv6_gso_segment_list_csum(skb); 476 else 477 return __udpv4_gso_segment_list_csum(skb); 478 } 479 480 struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb, 481 netdev_features_t features, bool is_ipv6) 482 { 483 struct sock *sk = gso_skb->sk; 484 unsigned int sum_truesize = 0; 485 struct sk_buff *segs, *seg; 486 __be16 newlen, msslen; 487 struct udphdr *uh; 488 unsigned int mss; 489 bool copy_dtor; 490 __sum16 check; 491 int ret = 0; 492 493 mss = skb_shinfo(gso_skb)->gso_size; 494 if (gso_skb->len <= sizeof(*uh) + mss) 495 return ERR_PTR(-EINVAL); 496 497 if (unlikely(skb_checksum_start(gso_skb) != 498 skb_transport_header(gso_skb) && 499 !(skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST))) 500 return ERR_PTR(-EINVAL); 501 502 /* We don't know if egress device can segment and checksum the packet 503 * when IPv6 extension headers are present. Fall back to software GSO. 504 */ 505 if (gso_skb->ip_summed != CHECKSUM_PARTIAL) 506 features &= ~(NETIF_F_GSO_UDP_L4 | NETIF_F_CSUM_MASK); 507 508 if (skb_gso_ok(gso_skb, features | NETIF_F_GSO_ROBUST)) { 509 /* Packet is from an untrusted source, reset gso_segs. */ 510 skb_shinfo(gso_skb)->gso_segs = DIV_ROUND_UP(gso_skb->len - sizeof(*uh), 511 mss); 512 return NULL; 513 } 514 515 if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST) { 516 /* Detect modified geometry and pass those to skb_segment. */ 517 if ((skb_pagelen(gso_skb) - sizeof(*uh) == skb_shinfo(gso_skb)->gso_size) && 518 !(skb_shinfo(gso_skb)->gso_type & SKB_GSO_DODGY)) 519 return __udp_gso_segment_list(gso_skb, features, is_ipv6); 520 521 ret = __skb_linearize(gso_skb); 522 if (ret) 523 return ERR_PTR(ret); 524 525 /* Setup csum, as fraglist skips this in udp4_gro_receive. */ 526 gso_skb->csum_start = skb_transport_header(gso_skb) - gso_skb->head; 527 gso_skb->csum_offset = offsetof(struct udphdr, check); 528 gso_skb->ip_summed = CHECKSUM_PARTIAL; 529 530 uh = udp_hdr(gso_skb); 531 if (is_ipv6) 532 uh->check = ~udp_v6_check(gso_skb->len, 533 &ipv6_hdr(gso_skb)->saddr, 534 &ipv6_hdr(gso_skb)->daddr, 0); 535 else 536 uh->check = ~udp_v4_check(gso_skb->len, 537 ip_hdr(gso_skb)->saddr, 538 ip_hdr(gso_skb)->daddr, 0); 539 } 540 541 skb_pull(gso_skb, sizeof(*uh)); 542 543 /* clear destructor to avoid skb_segment assigning it to tail */ 544 copy_dtor = gso_skb->destructor == sock_wfree; 545 if (copy_dtor) { 546 gso_skb->destructor = NULL; 547 gso_skb->sk = NULL; 548 } 549 550 segs = skb_segment(gso_skb, features); 551 if (IS_ERR_OR_NULL(segs)) { 552 if (copy_dtor) { 553 gso_skb->destructor = sock_wfree; 554 gso_skb->sk = sk; 555 } 556 return segs; 557 } 558 559 msslen = htons(sizeof(*uh) + mss); 560 561 /* GSO partial and frag_list segmentation only requires splitting 562 * the frame into an MSS multiple and possibly a remainder, both 563 * cases return a GSO skb. So update the mss now. 564 */ 565 if (skb_is_gso(segs)) 566 mss *= skb_shinfo(segs)->gso_segs; 567 568 seg = segs; 569 uh = udp_hdr(seg); 570 571 /* preserve TX timestamp flags and TS key for first segment */ 572 skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey; 573 skb_shinfo(seg)->tx_flags |= 574 (skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP); 575 576 /* compute checksum adjustment based on old length versus new */ 577 newlen = htons(sizeof(*uh) + mss); 578 check = csum16_add(csum16_sub(uh->check, uh->len), newlen); 579 580 for (;;) { 581 if (copy_dtor) { 582 seg->destructor = sock_wfree; 583 seg->sk = sk; 584 sum_truesize += seg->truesize; 585 } 586 587 if (!seg->next) 588 break; 589 590 uh->len = msslen; 591 uh->check = check; 592 593 if (seg->ip_summed == CHECKSUM_PARTIAL) 594 gso_reset_checksum(seg, ~check); 595 else 596 uh->check = gso_make_checksum(seg, ~check) ? : 597 CSUM_MANGLED_0; 598 599 seg = seg->next; 600 uh = udp_hdr(seg); 601 } 602 603 /* last packet can be partial gso_size, account for that in checksum */ 604 newlen = htons(skb_tail_pointer(seg) - skb_transport_header(seg) + 605 seg->data_len); 606 check = csum16_add(csum16_sub(uh->check, uh->len), newlen); 607 608 uh->len = newlen; 609 uh->check = check; 610 611 if (seg->ip_summed == CHECKSUM_PARTIAL) 612 gso_reset_checksum(seg, ~check); 613 else 614 uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0; 615 616 /* On the TX path, CHECKSUM_NONE and CHECKSUM_UNNECESSARY have the same 617 * meaning. However, check for bad offloads in the GSO stack expects the 618 * latter, if the checksum was calculated in software. To vouch for the 619 * segment skbs we actually need to set it on the gso_skb. 620 */ 621 if (gso_skb->ip_summed == CHECKSUM_NONE) 622 gso_skb->ip_summed = CHECKSUM_UNNECESSARY; 623 624 /* update refcount for the packet */ 625 if (copy_dtor) { 626 int delta = sum_truesize - gso_skb->truesize; 627 628 /* In some pathological cases, delta can be negative. 629 * We need to either use refcount_add() or refcount_sub_and_test() 630 */ 631 if (likely(delta >= 0)) 632 refcount_add(delta, &sk->sk_wmem_alloc); 633 else 634 WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc)); 635 } 636 return segs; 637 } 638 EXPORT_SYMBOL_GPL(__udp_gso_segment); 639 640 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb, 641 netdev_features_t features) 642 { 643 struct sk_buff *segs = ERR_PTR(-EINVAL); 644 unsigned int mss; 645 __wsum csum; 646 struct udphdr *uh; 647 struct iphdr *iph; 648 649 if (skb->encapsulation && 650 (skb_shinfo(skb)->gso_type & 651 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) { 652 segs = skb_udp_tunnel_segment(skb, features, false); 653 goto out; 654 } 655 656 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4))) 657 goto out; 658 659 if (!pskb_may_pull(skb, sizeof(struct udphdr))) 660 goto out; 661 662 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) 663 return __udp_gso_segment(skb, features, false); 664 665 mss = skb_shinfo(skb)->gso_size; 666 if (unlikely(skb->len <= mss)) 667 goto out; 668 669 /* Do software UFO. Complete and fill in the UDP checksum as 670 * HW cannot do checksum of UDP packets sent as multiple 671 * IP fragments. 672 */ 673 674 uh = udp_hdr(skb); 675 iph = ip_hdr(skb); 676 677 uh->check = 0; 678 csum = skb_checksum(skb, 0, skb->len, 0); 679 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum); 680 if (uh->check == 0) 681 uh->check = CSUM_MANGLED_0; 682 683 skb->ip_summed = CHECKSUM_UNNECESSARY; 684 685 /* If there is no outer header we can fake a checksum offload 686 * due to the fact that we have already done the checksum in 687 * software prior to segmenting the frame. 688 */ 689 if (!skb->encap_hdr_csum) 690 features |= NETIF_F_HW_CSUM; 691 692 /* Fragment the skb. IP headers of the fragments are updated in 693 * inet_gso_segment() 694 */ 695 segs = skb_segment(skb, features); 696 out: 697 return segs; 698 } 699 700 701 #define UDP_GRO_CNT_MAX 64 702 static struct sk_buff *udp_gro_receive_segment(struct list_head *head, 703 struct sk_buff *skb) 704 { 705 struct udphdr *uh = udp_gro_udphdr(skb); 706 struct sk_buff *pp = NULL; 707 struct udphdr *uh2; 708 struct sk_buff *p; 709 unsigned int ulen; 710 int ret = 0; 711 int flush; 712 713 /* requires non zero csum, for symmetry with GSO */ 714 if (!uh->check) { 715 NAPI_GRO_CB(skb)->flush = 1; 716 return NULL; 717 } 718 719 /* Do not deal with padded or malicious packets, sorry ! */ 720 ulen = ntohs(uh->len); 721 if (ulen <= sizeof(*uh) || ulen != skb_gro_len(skb)) { 722 NAPI_GRO_CB(skb)->flush = 1; 723 return NULL; 724 } 725 /* pull encapsulating udp header */ 726 skb_gro_pull(skb, sizeof(struct udphdr)); 727 728 list_for_each_entry(p, head, list) { 729 if (!NAPI_GRO_CB(p)->same_flow) 730 continue; 731 732 uh2 = udp_hdr(p); 733 734 /* Match ports only, as csum is always non zero */ 735 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) { 736 NAPI_GRO_CB(p)->same_flow = 0; 737 continue; 738 } 739 740 if (NAPI_GRO_CB(skb)->is_flist != NAPI_GRO_CB(p)->is_flist) { 741 NAPI_GRO_CB(skb)->flush = 1; 742 return p; 743 } 744 745 flush = gro_receive_network_flush(uh, uh2, p); 746 747 /* Terminate the flow on len mismatch or if it grow "too much". 748 * Under small packet flood GRO count could elsewhere grow a lot 749 * leading to excessive truesize values. 750 * On len mismatch merge the first packet shorter than gso_size, 751 * otherwise complete the GRO packet. 752 */ 753 if (ulen > ntohs(uh2->len) || flush) { 754 pp = p; 755 } else { 756 if (NAPI_GRO_CB(skb)->is_flist) { 757 if (!pskb_may_pull(skb, skb_gro_offset(skb))) { 758 NAPI_GRO_CB(skb)->flush = 1; 759 return NULL; 760 } 761 if ((skb->ip_summed != p->ip_summed) || 762 (skb->csum_level != p->csum_level)) { 763 NAPI_GRO_CB(skb)->flush = 1; 764 return NULL; 765 } 766 skb_set_network_header(skb, skb_gro_receive_network_offset(skb)); 767 ret = skb_gro_receive_list(p, skb); 768 } else { 769 skb_gro_postpull_rcsum(skb, uh, 770 sizeof(struct udphdr)); 771 772 ret = skb_gro_receive(p, skb); 773 } 774 } 775 776 if (ret || ulen != ntohs(uh2->len) || 777 NAPI_GRO_CB(p)->count >= UDP_GRO_CNT_MAX) 778 pp = p; 779 780 return pp; 781 } 782 783 /* mismatch, but we never need to flush */ 784 return NULL; 785 } 786 787 struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb, 788 struct udphdr *uh, struct sock *sk) 789 { 790 struct sk_buff *pp = NULL; 791 struct sk_buff *p; 792 struct udphdr *uh2; 793 unsigned int off = skb_gro_offset(skb); 794 int flush = 1; 795 796 /* We can do L4 aggregation only if the packet can't land in a tunnel 797 * otherwise we could corrupt the inner stream. Detecting such packets 798 * cannot be foolproof and the aggregation might still happen in some 799 * cases. Such packets should be caught in udp_unexpected_gso later. 800 */ 801 NAPI_GRO_CB(skb)->is_flist = 0; 802 if (!sk || !udp_sk(sk)->gro_receive) { 803 /* If the packet was locally encapsulated in a UDP tunnel that 804 * wasn't detected above, do not GRO. 805 */ 806 if (skb->encapsulation) 807 goto out; 808 809 if (skb->dev->features & NETIF_F_GRO_FRAGLIST) 810 NAPI_GRO_CB(skb)->is_flist = sk ? !udp_test_bit(GRO_ENABLED, sk) : 1; 811 812 if ((!sk && (skb->dev->features & NETIF_F_GRO_UDP_FWD)) || 813 (sk && udp_test_bit(GRO_ENABLED, sk)) || NAPI_GRO_CB(skb)->is_flist) 814 return call_gro_receive(udp_gro_receive_segment, head, skb); 815 816 /* no GRO, be sure flush the current packet */ 817 goto out; 818 } 819 820 if (NAPI_GRO_CB(skb)->encap_mark || 821 (uh->check && skb->ip_summed != CHECKSUM_PARTIAL && 822 NAPI_GRO_CB(skb)->csum_cnt == 0 && 823 !NAPI_GRO_CB(skb)->csum_valid)) 824 goto out; 825 826 /* mark that this skb passed once through the tunnel gro layer */ 827 NAPI_GRO_CB(skb)->encap_mark = 1; 828 829 flush = 0; 830 831 list_for_each_entry(p, head, list) { 832 if (!NAPI_GRO_CB(p)->same_flow) 833 continue; 834 835 uh2 = (struct udphdr *)(p->data + off); 836 837 /* Match ports and either checksums are either both zero 838 * or nonzero. 839 */ 840 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) || 841 (!uh->check ^ !uh2->check)) { 842 NAPI_GRO_CB(p)->same_flow = 0; 843 continue; 844 } 845 } 846 847 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */ 848 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr)); 849 pp = udp_tunnel_gro_rcv(sk, head, skb); 850 851 out: 852 skb_gro_flush_final(skb, pp, flush); 853 return pp; 854 } 855 EXPORT_SYMBOL(udp_gro_receive); 856 857 static struct sock *udp4_gro_lookup_skb(struct sk_buff *skb, __be16 sport, 858 __be16 dport) 859 { 860 const struct iphdr *iph = skb_gro_network_header(skb); 861 struct net *net = dev_net_rcu(skb->dev); 862 struct sock *sk; 863 int iif, sdif; 864 865 sk = udp_tunnel_sk(net, false); 866 if (sk && dport == htons(sk->sk_num)) 867 return sk; 868 869 inet_get_iif_sdif(skb, &iif, &sdif); 870 871 return __udp4_lib_lookup(net, iph->saddr, sport, 872 iph->daddr, dport, iif, 873 sdif, net->ipv4.udp_table, NULL); 874 } 875 876 INDIRECT_CALLABLE_SCOPE 877 struct sk_buff *udp4_gro_receive(struct list_head *head, struct sk_buff *skb) 878 { 879 struct udphdr *uh = udp_gro_udphdr(skb); 880 struct sock *sk = NULL; 881 struct sk_buff *pp; 882 883 if (unlikely(!uh)) 884 goto flush; 885 886 /* Don't bother verifying checksum if we're going to flush anyway. */ 887 if (NAPI_GRO_CB(skb)->flush) 888 goto skip; 889 890 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check, 891 inet_gro_compute_pseudo)) 892 goto flush; 893 else if (uh->check) 894 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, 895 inet_gro_compute_pseudo); 896 skip: 897 if (static_branch_unlikely(&udp_encap_needed_key)) 898 sk = udp4_gro_lookup_skb(skb, uh->source, uh->dest); 899 900 pp = udp_gro_receive(head, skb, uh, sk); 901 return pp; 902 903 flush: 904 NAPI_GRO_CB(skb)->flush = 1; 905 return NULL; 906 } 907 908 static int udp_gro_complete_segment(struct sk_buff *skb) 909 { 910 struct udphdr *uh = udp_hdr(skb); 911 912 skb->csum_start = (unsigned char *)uh - skb->head; 913 skb->csum_offset = offsetof(struct udphdr, check); 914 skb->ip_summed = CHECKSUM_PARTIAL; 915 916 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count; 917 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_L4; 918 919 if (skb->encapsulation) 920 skb->inner_transport_header = skb->transport_header; 921 922 return 0; 923 } 924 925 int udp_gro_complete(struct sk_buff *skb, int nhoff, 926 udp_lookup_t lookup) 927 { 928 __be16 newlen = htons(skb->len - nhoff); 929 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff); 930 struct sock *sk; 931 int err; 932 933 uh->len = newlen; 934 935 sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb, 936 udp4_lib_lookup_skb, skb, uh->source, uh->dest); 937 if (sk && udp_sk(sk)->gro_complete) { 938 skb_shinfo(skb)->gso_type = uh->check ? SKB_GSO_UDP_TUNNEL_CSUM 939 : SKB_GSO_UDP_TUNNEL; 940 941 /* clear the encap mark, so that inner frag_list gro_complete 942 * can take place 943 */ 944 NAPI_GRO_CB(skb)->encap_mark = 0; 945 946 /* Set encapsulation before calling into inner gro_complete() 947 * functions to make them set up the inner offsets. 948 */ 949 skb->encapsulation = 1; 950 err = udp_sk(sk)->gro_complete(sk, skb, 951 nhoff + sizeof(struct udphdr)); 952 } else { 953 err = udp_gro_complete_segment(skb); 954 } 955 956 if (skb->remcsum_offload) 957 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM; 958 959 return err; 960 } 961 EXPORT_SYMBOL(udp_gro_complete); 962 963 INDIRECT_CALLABLE_SCOPE int udp4_gro_complete(struct sk_buff *skb, int nhoff) 964 { 965 const u16 offset = NAPI_GRO_CB(skb)->network_offsets[skb->encapsulation]; 966 const struct iphdr *iph = (struct iphdr *)(skb->data + offset); 967 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff); 968 969 /* do fraglist only if there is no outer UDP encap (or we already processed it) */ 970 if (NAPI_GRO_CB(skb)->is_flist && !NAPI_GRO_CB(skb)->encap_mark) { 971 uh->len = htons(skb->len - nhoff); 972 973 skb_shinfo(skb)->gso_type |= (SKB_GSO_FRAGLIST|SKB_GSO_UDP_L4); 974 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count; 975 976 __skb_incr_checksum_unnecessary(skb); 977 978 return 0; 979 } 980 981 if (uh->check) 982 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr, 983 iph->daddr, 0); 984 985 return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb); 986 } 987 988 int __init udpv4_offload_init(void) 989 { 990 net_hotdata.udpv4_offload = (struct net_offload) { 991 .callbacks = { 992 .gso_segment = udp4_ufo_fragment, 993 .gro_receive = udp4_gro_receive, 994 .gro_complete = udp4_gro_complete, 995 }, 996 }; 997 998 return inet_add_offload(&net_hotdata.udpv4_offload, IPPROTO_UDP); 999 } 1000