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         /* preserve TX timestamp flags and TS key for first segment */
 231         skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey;
 232         skb_shinfo(seg)->tx_flags |=
 233                         (skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP);
 234
 235         /* compute checksum adjustment based on old length versus new */
 236         newlen = htons(sizeof(*uh) + mss);
 237         check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
 238
 239         for (;;) {
 240                 if (copy_dtor) {
 241                         seg->destructor = sock_wfree;
 242                         seg->sk = sk;
 243                         sum_truesize += seg->truesize;
 244                 }
 245
 246                 if (!seg->next)
 247                         break;
 248
 249                 uh->len = newlen;
 250                 uh->check = check;
 251
 252                 if (seg->ip_summed == CHECKSUM_PARTIAL)
 253                         gso_reset_checksum(seg, ~check);
 254                 else
 255                         uh->check = gso_make_checksum(seg, ~check) ? :
 256                                     CSUM_MANGLED_0;
 257
 258                 seg = seg->next;
 259                 uh = udp_hdr(seg);
 260         }
 261
 262         /* last packet can be partial gso_size, account for that in checksum */
 263         newlen = htons(skb_tail_pointer(seg) - skb_transport_header(seg) +
 264                        seg->data_len);
 265         check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
 266
 267         uh->len = newlen;
 268         uh->check = check;
 269
 270         if (seg->ip_summed == CHECKSUM_PARTIAL)
 271                 gso_reset_checksum(seg, ~check);
 272         else
 273                 uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0;
 274
 275         /* update refcount for the packet */
 276         if (copy_dtor) {
 277                 int delta = sum_truesize - gso_skb->truesize;
 278
 279                 /* In some pathological cases, delta can be negative.
 280                  * We need to either use refcount_add() or refcount_sub_and_test()
 281                  */
 282                 if (likely(delta >= 0))
 283                         refcount_add(delta, &sk->sk_wmem_alloc);
 284                 else
 285                         WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));
 286         }
 287         return segs;
 288 }
 289 EXPORT_SYMBOL_GPL(__udp_gso_segment);
 290
 291 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
 292                                          netdev_features_t features)
 293 {
 294         struct sk_buff *segs = ERR_PTR(-EINVAL);
 295         unsigned int mss;
 296         __wsum csum;
 297         struct udphdr *uh;
 298         struct iphdr *iph;
 299
 300         if (skb->encapsulation &&
 301             (skb_shinfo(skb)->gso_type &
 302              (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
 303                 segs = skb_udp_tunnel_segment(skb, features, false);
 304                 goto out;
 305         }
 306
 307         if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4)))
 308                 goto out;
 309
 310         if (!pskb_may_pull(skb, sizeof(struct udphdr)))
 311                 goto out;
 312
 313         if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
 314                 return __udp_gso_segment(skb, features);
 315
 316         mss = skb_shinfo(skb)->gso_size;
 317         if (unlikely(skb->len <= mss))
 318                 goto out;
 319
 320         /* Do software UFO. Complete and fill in the UDP checksum as
 321          * HW cannot do checksum of UDP packets sent as multiple
 322          * IP fragments.
 323          */
 324
 325         uh = udp_hdr(skb);
 326         iph = ip_hdr(skb);
 327
 328         uh->check = 0;
 329         csum = skb_checksum(skb, 0, skb->len, 0);
 330         uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
 331         if (uh->check == 0)
 332                 uh->check = CSUM_MANGLED_0;
 333
 334         skb->ip_summed = CHECKSUM_UNNECESSARY;
 335
 336         /* If there is no outer header we can fake a checksum offload
 337          * due to the fact that we have already done the checksum in
 338          * software prior to segmenting the frame.
 339          */
 340         if (!skb->encap_hdr_csum)
 341                 features |= NETIF_F_HW_CSUM;
 342
 343         /* Fragment the skb. IP headers of the fragments are updated in
 344          * inet_gso_segment()
 345          */
 346         segs = skb_segment(skb, features);
 347 out:
 348         return segs;
 349 }
 350
 351 struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb,
 352                                 struct udphdr *uh, udp_lookup_t lookup)
 353 {
 354         struct sk_buff *pp = NULL;
 355         struct sk_buff *p;
 356         struct udphdr *uh2;
 357         unsigned int off = skb_gro_offset(skb);
 358         int flush = 1;
 359         struct sock *sk;
 360
 361         if (NAPI_GRO_CB(skb)->encap_mark ||
 362             (uh->check && skb->ip_summed != CHECKSUM_PARTIAL &&
 363              NAPI_GRO_CB(skb)->csum_cnt == 0 &&
 364              !NAPI_GRO_CB(skb)->csum_valid))
 365                 goto out;
 366
 367         /* mark that this skb passed once through the tunnel gro layer */
 368         NAPI_GRO_CB(skb)->encap_mark = 1;
 369
 370         rcu_read_lock();
 371         sk = (*lookup)(skb, uh->source, uh->dest);
 372
 373         if (sk && udp_sk(sk)->gro_receive)
 374                 goto unflush;
 375         goto out_unlock;
 376
 377 unflush:
 378         flush = 0;
 379
 380         list_for_each_entry(p, head, list) {
 381                 if (!NAPI_GRO_CB(p)->same_flow)
 382                         continue;
 383
 384                 uh2 = (struct udphdr   *)(p->data + off);
 385
 386                 /* Match ports and either checksums are either both zero
 387                  * or nonzero.
 388                  */
 389                 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
 390                     (!uh->check ^ !uh2->check)) {
 391                         NAPI_GRO_CB(p)->same_flow = 0;
 392                         continue;
 393                 }
 394         }
 395
 396         skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
 397         skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
 398         pp = call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
 399
 400 out_unlock:
 401         rcu_read_unlock();
 402 out:
 403         skb_gro_flush_final(skb, pp, flush);
 404         return pp;
 405 }
 406 EXPORT_SYMBOL(udp_gro_receive);
 407
 408 static struct sk_buff *udp4_gro_receive(struct list_head *head,
 409                                         struct sk_buff *skb)
 410 {
 411         struct udphdr *uh = udp_gro_udphdr(skb);
 412
 413         if (unlikely(!uh))
 414                 goto flush;
 415
 416         /* Don't bother verifying checksum if we're going to flush anyway. */
 417         if (NAPI_GRO_CB(skb)->flush)
 418                 goto skip;
 419
 420         if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
 421                                                  inet_gro_compute_pseudo))
 422                 goto flush;
 423         else if (uh->check)
 424                 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, uh->check,
 425                                              inet_gro_compute_pseudo);
 426 skip:
 427         NAPI_GRO_CB(skb)->is_ipv6 = 0;
 428         return udp_gro_receive(head, skb, uh, udp4_lib_lookup_skb);
 429
 430 flush:
 431         NAPI_GRO_CB(skb)->flush = 1;
 432         return NULL;
 433 }
 434
 435 int udp_gro_complete(struct sk_buff *skb, int nhoff,
 436                      udp_lookup_t lookup)
 437 {
 438         __be16 newlen = htons(skb->len - nhoff);
 439         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 440         int err = -ENOSYS;
 441         struct sock *sk;
 442
 443         uh->len = newlen;
 444
 445         /* Set encapsulation before calling into inner gro_complete() functions
 446          * to make them set up the inner offsets.
 447          */
 448         skb->encapsulation = 1;
 449
 450         rcu_read_lock();
 451         sk = (*lookup)(skb, uh->source, uh->dest);
 452         if (sk && udp_sk(sk)->gro_complete)
 453                 err = udp_sk(sk)->gro_complete(sk, skb,
 454                                 nhoff + sizeof(struct udphdr));
 455         rcu_read_unlock();
 456
 457         if (skb->remcsum_offload)
 458                 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
 459
 460         return err;
 461 }
 462 EXPORT_SYMBOL(udp_gro_complete);
 463
 464 static int udp4_gro_complete(struct sk_buff *skb, int nhoff)
 465 {
 466         const struct iphdr *iph = ip_hdr(skb);
 467         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 468
 469         if (uh->check) {
 470                 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL_CSUM;
 471                 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
 472                                           iph->daddr, 0);
 473         } else {
 474                 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL;
 475         }
 476
 477         return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
 478 }
 479
 480 static const struct net_offload udpv4_offload = {
 481         .callbacks = {
 482                 .gso_segment = udp4_ufo_fragment,
 483                 .gro_receive  = udp4_gro_receive,
 484                 .gro_complete = udp4_gro_complete,
 485         },
 486 };
 487
 488 int __init udpv4_offload_init(void)
 489 {
 490         return inet_add_offload(&udpv4_offload, IPPROTO_UDP);
 491 }