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, ufo;
  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
  33         if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
  34                 goto out;
  35
  36         /* Adjust partial header checksum to negate old length.
  37          * We cannot rely on the value contained in uh->len as it is
  38          * possible that the actual value exceeds the boundaries of the
  39          * 16 bit length field due to the header being added outside of an
  40          * IP or IPv6 frame that was already limited to 64K - 1.
  41          */
  42         if (skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)
  43                 partial = (__force __wsum)uh->len;
  44         else
  45                 partial = (__force __wsum)htonl(skb->len);
  46         partial = csum_sub(csum_unfold(uh->check), partial);
  47
  48         /* setup inner skb. */
  49         skb->encapsulation = 0;
  50         SKB_GSO_CB(skb)->encap_level = 0;
  51         __skb_pull(skb, tnl_hlen);
  52         skb_reset_mac_header(skb);
  53         skb_set_network_header(skb, skb_inner_network_offset(skb));
  54         skb->mac_len = skb_inner_network_offset(skb);
  55         skb->protocol = new_protocol;
  56
  57         need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
  58         skb->encap_hdr_csum = need_csum;
  59
  60         remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
  61         skb->remcsum_offload = remcsum;
  62
  63         ufo = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP);
  64
  65         /* Try to offload checksum if possible */
  66         offload_csum = !!(need_csum &&
  67                           (skb->dev->features &
  68                            (is_ipv6 ? (NETIF_F_HW_CSUM | NETIF_F_IPV6_CSUM) :
  69                                       (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM))));
  70
  71         features &= skb->dev->hw_enc_features;
  72
  73         /* The only checksum offload we care about from here on out is the
  74          * outer one so strip the existing checksum feature flags and
  75          * instead set the flag based on our outer checksum offload value.
  76          */
  77         if (remcsum || ufo) {
  78                 features &= ~NETIF_F_CSUM_MASK;
  79                 if (!need_csum || offload_csum)
  80                         features |= NETIF_F_HW_CSUM;
  81         }
  82
  83         /* segment inner packet. */
  84         segs = gso_inner_segment(skb, features);
  85         if (IS_ERR_OR_NULL(segs)) {
  86                 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
  87                                      mac_len);
  88                 goto out;
  89         }
  90
  91         outer_hlen = skb_tnl_header_len(skb);
  92         udp_offset = outer_hlen - tnl_hlen;
  93         skb = segs;
  94         do {
  95                 unsigned int len;
  96
  97                 if (remcsum)
  98                         skb->ip_summed = CHECKSUM_NONE;
  99
 100                 /* Set up inner headers if we are offloading inner checksum */
 101                 if (skb->ip_summed == CHECKSUM_PARTIAL) {
 102                         skb_reset_inner_headers(skb);
 103                         skb->encapsulation = 1;
 104                 }
 105
 106                 skb->mac_len = mac_len;
 107                 skb->protocol = protocol;
 108
 109                 __skb_push(skb, outer_hlen);
 110                 skb_reset_mac_header(skb);
 111                 skb_set_network_header(skb, mac_len);
 112                 skb_set_transport_header(skb, udp_offset);
 113                 len = skb->len - udp_offset;
 114                 uh = udp_hdr(skb);
 115
 116                 /* If we are only performing partial GSO the inner header
 117                  * will be using a length value equal to only one MSS sized
 118                  * segment instead of the entire frame.
 119                  */
 120                 if (skb_is_gso(skb)) {
 121                         uh->len = htons(skb_shinfo(skb)->gso_size +
 122                                         SKB_GSO_CB(skb)->data_offset +
 123                                         skb->head - (unsigned char *)uh);
 124                 } else {
 125                         uh->len = htons(len);
 126                 }
 127
 128                 if (!need_csum)
 129                         continue;
 130
 131                 uh->check = ~csum_fold(csum_add(partial,
 132                                        (__force __wsum)htonl(len)));
 133
 134                 if (skb->encapsulation || !offload_csum) {
 135                         uh->check = gso_make_checksum(skb, ~uh->check);
 136                         if (uh->check == 0)
 137                                 uh->check = CSUM_MANGLED_0;
 138                 } else {
 139                         skb->ip_summed = CHECKSUM_PARTIAL;
 140                         skb->csum_start = skb_transport_header(skb) - skb->head;
 141                         skb->csum_offset = offsetof(struct udphdr, check);
 142                 }
 143         } while ((skb = skb->next));
 144 out:
 145         return segs;
 146 }
 147
 148 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
 149                                        netdev_features_t features,
 150                                        bool is_ipv6)
 151 {
 152         __be16 protocol = skb->protocol;
 153         const struct net_offload **offloads;
 154         const struct net_offload *ops;
 155         struct sk_buff *segs = ERR_PTR(-EINVAL);
 156         struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
 157                                              netdev_features_t features);
 158
 159         rcu_read_lock();
 160
 161         switch (skb->inner_protocol_type) {
 162         case ENCAP_TYPE_ETHER:
 163                 protocol = skb->inner_protocol;
 164                 gso_inner_segment = skb_mac_gso_segment;
 165                 break;
 166         case ENCAP_TYPE_IPPROTO:
 167                 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
 168                 ops = rcu_dereference(offloads[skb->inner_ipproto]);
 169                 if (!ops || !ops->callbacks.gso_segment)
 170                         goto out_unlock;
 171                 gso_inner_segment = ops->callbacks.gso_segment;
 172                 break;
 173         default:
 174                 goto out_unlock;
 175         }
 176
 177         segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
 178                                         protocol, is_ipv6);
 179
 180 out_unlock:
 181         rcu_read_unlock();
 182
 183         return segs;
 184 }
 185 EXPORT_SYMBOL(skb_udp_tunnel_segment);
 186
 187 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
 188                                          netdev_features_t features)
 189 {
 190         struct sk_buff *segs = ERR_PTR(-EINVAL);
 191         unsigned int mss;
 192         __wsum csum;
 193         struct udphdr *uh;
 194         struct iphdr *iph;
 195
 196         if (skb->encapsulation &&
 197             (skb_shinfo(skb)->gso_type &
 198              (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
 199                 segs = skb_udp_tunnel_segment(skb, features, false);
 200                 goto out;
 201         }
 202
 203         if (!pskb_may_pull(skb, sizeof(struct udphdr)))
 204                 goto out;
 205
 206         mss = skb_shinfo(skb)->gso_size;
 207         if (unlikely(skb->len <= mss))
 208                 goto out;
 209
 210         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
 211                 /* Packet is from an untrusted source, reset gso_segs. */
 212                 int type = skb_shinfo(skb)->gso_type;
 213
 214                 if (unlikely(type & ~(SKB_GSO_UDP | SKB_GSO_DODGY |
 215                                       SKB_GSO_UDP_TUNNEL |
 216                                       SKB_GSO_UDP_TUNNEL_CSUM |
 217                                       SKB_GSO_TUNNEL_REMCSUM |
 218                                       SKB_GSO_IPIP |
 219                                       SKB_GSO_GRE | SKB_GSO_GRE_CSUM) ||
 220                              !(type & (SKB_GSO_UDP))))
 221                         goto out;
 222
 223                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
 224
 225                 segs = NULL;
 226                 goto out;
 227         }
 228
 229         /* Do software UFO. Complete and fill in the UDP checksum as
 230          * HW cannot do checksum of UDP packets sent as multiple
 231          * IP fragments.
 232          */
 233
 234         uh = udp_hdr(skb);
 235         iph = ip_hdr(skb);
 236
 237         uh->check = 0;
 238         csum = skb_checksum(skb, 0, skb->len, 0);
 239         uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
 240         if (uh->check == 0)
 241                 uh->check = CSUM_MANGLED_0;
 242
 243         skb->ip_summed = CHECKSUM_NONE;
 244
 245         /* If there is no outer header we can fake a checksum offload
 246          * due to the fact that we have already done the checksum in
 247          * software prior to segmenting the frame.
 248          */
 249         if (!skb->encap_hdr_csum)
 250                 features |= NETIF_F_HW_CSUM;
 251
 252         /* Fragment the skb. IP headers of the fragments are updated in
 253          * inet_gso_segment()
 254          */
 255         segs = skb_segment(skb, features);
 256 out:
 257         return segs;
 258 }
 259
 260 struct sk_buff **udp_gro_receive(struct sk_buff **head, struct sk_buff *skb,
 261                                  struct udphdr *uh, udp_lookup_t lookup)
 262 {
 263         struct sk_buff *p, **pp = NULL;
 264         struct udphdr *uh2;
 265         unsigned int off = skb_gro_offset(skb);
 266         int flush = 1;
 267         struct sock *sk;
 268
 269         if (NAPI_GRO_CB(skb)->encap_mark ||
 270             (skb->ip_summed != CHECKSUM_PARTIAL &&
 271              NAPI_GRO_CB(skb)->csum_cnt == 0 &&
 272              !NAPI_GRO_CB(skb)->csum_valid))
 273                 goto out;
 274
 275         /* mark that this skb passed once through the tunnel gro layer */
 276         NAPI_GRO_CB(skb)->encap_mark = 1;
 277
 278         rcu_read_lock();
 279         sk = (*lookup)(skb, uh->source, uh->dest);
 280
 281         if (sk && udp_sk(sk)->gro_receive)
 282                 goto unflush;
 283         goto out_unlock;
 284
 285 unflush:
 286         flush = 0;
 287
 288         for (p = *head; p; p = p->next) {
 289                 if (!NAPI_GRO_CB(p)->same_flow)
 290                         continue;
 291
 292                 uh2 = (struct udphdr   *)(p->data + off);
 293
 294                 /* Match ports and either checksums are either both zero
 295                  * or nonzero.
 296                  */
 297                 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
 298                     (!uh->check ^ !uh2->check)) {
 299                         NAPI_GRO_CB(p)->same_flow = 0;
 300                         continue;
 301                 }
 302         }
 303
 304         skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
 305         skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
 306         pp = udp_sk(sk)->gro_receive(sk, head, skb);
 307
 308 out_unlock:
 309         rcu_read_unlock();
 310 out:
 311         NAPI_GRO_CB(skb)->flush |= flush;
 312         return pp;
 313 }
 314 EXPORT_SYMBOL(udp_gro_receive);
 315
 316 static struct sk_buff **udp4_gro_receive(struct sk_buff **head,
 317                                          struct sk_buff *skb)
 318 {
 319         struct udphdr *uh = udp_gro_udphdr(skb);
 320
 321         if (unlikely(!uh))
 322                 goto flush;
 323
 324         /* Don't bother verifying checksum if we're going to flush anyway. */
 325         if (NAPI_GRO_CB(skb)->flush)
 326                 goto skip;
 327
 328         if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
 329                                                  inet_gro_compute_pseudo))
 330                 goto flush;
 331         else if (uh->check)
 332                 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, uh->check,
 333                                              inet_gro_compute_pseudo);
 334 skip:
 335         NAPI_GRO_CB(skb)->is_ipv6 = 0;
 336         return udp_gro_receive(head, skb, uh, udp4_lib_lookup_skb);
 337
 338 flush:
 339         NAPI_GRO_CB(skb)->flush = 1;
 340         return NULL;
 341 }
 342
 343 int udp_gro_complete(struct sk_buff *skb, int nhoff,
 344                      udp_lookup_t lookup)
 345 {
 346         __be16 newlen = htons(skb->len - nhoff);
 347         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 348         int err = -ENOSYS;
 349         struct sock *sk;
 350
 351         uh->len = newlen;
 352
 353         /* Set encapsulation before calling into inner gro_complete() functions
 354          * to make them set up the inner offsets.
 355          */
 356         skb->encapsulation = 1;
 357
 358         rcu_read_lock();
 359         sk = (*lookup)(skb, uh->source, uh->dest);
 360         if (sk && udp_sk(sk)->gro_complete)
 361                 err = udp_sk(sk)->gro_complete(sk, skb,
 362                                 nhoff + sizeof(struct udphdr));
 363         rcu_read_unlock();
 364
 365         if (skb->remcsum_offload)
 366                 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
 367
 368         return err;
 369 }
 370 EXPORT_SYMBOL(udp_gro_complete);
 371
 372 static int udp4_gro_complete(struct sk_buff *skb, int nhoff)
 373 {
 374         const struct iphdr *iph = ip_hdr(skb);
 375         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 376
 377         if (uh->check) {
 378                 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL_CSUM;
 379                 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
 380                                           iph->daddr, 0);
 381         } else {
 382                 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL;
 383         }
 384
 385         return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
 386 }
 387
 388 static const struct net_offload udpv4_offload = {
 389         .callbacks = {
 390                 .gso_segment = udp4_ufo_fragment,
 391                 .gro_receive  = udp4_gro_receive,
 392                 .gro_complete = udp4_gro_complete,
 393         },
 394 };
 395
 396 int __init udpv4_offload_init(void)
 397 {
 398         return inet_add_offload(&udpv4_offload, IPPROTO_UDP);
 399 }