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