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