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