net/ipv4/udp_offload.c

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