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 DEFINE_SPINLOCK(udp_offload_lock);
  18 static struct udp_offload_priv __rcu *udp_offload_base __read_mostly;
  19
  20 #define udp_deref_protected(X) rcu_dereference_protected(X, lockdep_is_held(&udp_offload_lock))
  21
  22 struct udp_offload_priv {
  23         struct udp_offload      *offload;
  24         possible_net_t  net;
  25         struct rcu_head         rcu;
  26         struct udp_offload_priv __rcu *next;
  27 };
  28
  29 static struct sk_buff *__skb_udp_tunnel_segment(struct sk_buff *skb,
  30         netdev_features_t features,
  31         struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
  32                                              netdev_features_t features),
  33         __be16 new_protocol, bool is_ipv6)
  34 {
  35         int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
  36         bool remcsum, need_csum, offload_csum, ufo;
  37         struct sk_buff *segs = ERR_PTR(-EINVAL);
  38         struct udphdr *uh = udp_hdr(skb);
  39         u16 mac_offset = skb->mac_header;
  40         __be16 protocol = skb->protocol;
  41         u16 mac_len = skb->mac_len;
  42         int udp_offset, outer_hlen;
  43         __wsum partial;
  44
  45         if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
  46                 goto out;
  47
  48         /* Adjust partial header checksum to negate old length.
  49          * We cannot rely on the value contained in uh->len as it is
  50          * possible that the actual value exceeds the boundaries of the
  51          * 16 bit length field due to the header being added outside of an
  52          * IP or IPv6 frame that was already limited to 64K - 1.
  53          */
  54         partial = csum_sub(csum_unfold(uh->check),
  55                            (__force __wsum)htonl(skb->len));
  56
  57         /* setup inner skb. */
  58         skb->encapsulation = 0;
  59         SKB_GSO_CB(skb)->encap_level = 0;
  60         __skb_pull(skb, tnl_hlen);
  61         skb_reset_mac_header(skb);
  62         skb_set_network_header(skb, skb_inner_network_offset(skb));
  63         skb->mac_len = skb_inner_network_offset(skb);
  64         skb->protocol = new_protocol;
  65
  66         need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
  67         skb->encap_hdr_csum = need_csum;
  68
  69         remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
  70         skb->remcsum_offload = remcsum;
  71
  72         ufo = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP);
  73
  74         /* Try to offload checksum if possible */
  75         offload_csum = !!(need_csum &&
  76                           (skb->dev->features &
  77                            (is_ipv6 ? (NETIF_F_HW_CSUM | NETIF_F_IPV6_CSUM) :
  78                                       (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM))));
  79
  80         features &= skb->dev->hw_enc_features;
  81
  82         /* The only checksum offload we care about from here on out is the
  83          * outer one so strip the existing checksum feature flags and
  84          * instead set the flag based on our outer checksum offload value.
  85          */
  86         if (remcsum || ufo) {
  87                 features &= ~NETIF_F_CSUM_MASK;
  88                 if (!need_csum || offload_csum)
  89                         features |= NETIF_F_HW_CSUM;
  90         }
  91
  92         /* segment inner packet. */
  93         segs = gso_inner_segment(skb, features);
  94         if (IS_ERR_OR_NULL(segs)) {
  95                 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
  96                                      mac_len);
  97                 goto out;
  98         }
  99
 100         outer_hlen = skb_tnl_header_len(skb);
 101         udp_offset = outer_hlen - tnl_hlen;
 102         skb = segs;
 103         do {
 104                 __be16 len;
 105
 106                 if (remcsum)
 107                         skb->ip_summed = CHECKSUM_NONE;
 108
 109                 /* Set up inner headers if we are offloading inner checksum */
 110                 if (skb->ip_summed == CHECKSUM_PARTIAL) {
 111                         skb_reset_inner_headers(skb);
 112                         skb->encapsulation = 1;
 113                 }
 114
 115                 skb->mac_len = mac_len;
 116                 skb->protocol = protocol;
 117
 118                 __skb_push(skb, outer_hlen);
 119                 skb_reset_mac_header(skb);
 120                 skb_set_network_header(skb, mac_len);
 121                 skb_set_transport_header(skb, udp_offset);
 122                 len = htons(skb->len - udp_offset);
 123                 uh = udp_hdr(skb);
 124                 uh->len = len;
 125
 126                 if (!need_csum)
 127                         continue;
 128
 129                 uh->check = ~csum_fold(csum_add(partial, (__force __wsum)len));
 130
 131                 if (skb->encapsulation || !offload_csum) {
 132                         uh->check = gso_make_checksum(skb, ~uh->check);
 133                         if (uh->check == 0)
 134                                 uh->check = CSUM_MANGLED_0;
 135                 } else {
 136                         skb->ip_summed = CHECKSUM_PARTIAL;
 137                         skb->csum_start = skb_transport_header(skb) - skb->head;
 138                         skb->csum_offset = offsetof(struct udphdr, check);
 139                 }
 140         } while ((skb = skb->next));
 141 out:
 142         return segs;
 143 }
 144
 145 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
 146                                        netdev_features_t features,
 147                                        bool is_ipv6)
 148 {
 149         __be16 protocol = skb->protocol;
 150         const struct net_offload **offloads;
 151         const struct net_offload *ops;
 152         struct sk_buff *segs = ERR_PTR(-EINVAL);
 153         struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
 154                                              netdev_features_t features);
 155
 156         rcu_read_lock();
 157
 158         switch (skb->inner_protocol_type) {
 159         case ENCAP_TYPE_ETHER:
 160                 protocol = skb->inner_protocol;
 161                 gso_inner_segment = skb_mac_gso_segment;
 162                 break;
 163         case ENCAP_TYPE_IPPROTO:
 164                 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
 165                 ops = rcu_dereference(offloads[skb->inner_ipproto]);
 166                 if (!ops || !ops->callbacks.gso_segment)
 167                         goto out_unlock;
 168                 gso_inner_segment = ops->callbacks.gso_segment;
 169                 break;
 170         default:
 171                 goto out_unlock;
 172         }
 173
 174         segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
 175                                         protocol, is_ipv6);
 176
 177 out_unlock:
 178         rcu_read_unlock();
 179
 180         return segs;
 181 }
 182
 183 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
 184                                          netdev_features_t features)
 185 {
 186         struct sk_buff *segs = ERR_PTR(-EINVAL);
 187         unsigned int mss;
 188         __wsum csum;
 189         struct udphdr *uh;
 190         struct iphdr *iph;
 191
 192         if (skb->encapsulation &&
 193             (skb_shinfo(skb)->gso_type &
 194              (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
 195                 segs = skb_udp_tunnel_segment(skb, features, false);
 196                 goto out;
 197         }
 198
 199         if (!pskb_may_pull(skb, sizeof(struct udphdr)))
 200                 goto out;
 201
 202         mss = skb_shinfo(skb)->gso_size;
 203         if (unlikely(skb->len <= mss))
 204                 goto out;
 205
 206         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
 207                 /* Packet is from an untrusted source, reset gso_segs. */
 208                 int type = skb_shinfo(skb)->gso_type;
 209
 210                 if (unlikely(type & ~(SKB_GSO_UDP | SKB_GSO_DODGY |
 211                                       SKB_GSO_UDP_TUNNEL |
 212                                       SKB_GSO_UDP_TUNNEL_CSUM |
 213                                       SKB_GSO_TUNNEL_REMCSUM |
 214                                       SKB_GSO_IPIP |
 215                                       SKB_GSO_GRE | SKB_GSO_GRE_CSUM) ||
 216                              !(type & (SKB_GSO_UDP))))
 217                         goto out;
 218
 219                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
 220
 221                 segs = NULL;
 222                 goto out;
 223         }
 224
 225         /* Do software UFO. Complete and fill in the UDP checksum as
 226          * HW cannot do checksum of UDP packets sent as multiple
 227          * IP fragments.
 228          */
 229
 230         uh = udp_hdr(skb);
 231         iph = ip_hdr(skb);
 232
 233         uh->check = 0;
 234         csum = skb_checksum(skb, 0, skb->len, 0);
 235         uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
 236         if (uh->check == 0)
 237                 uh->check = CSUM_MANGLED_0;
 238
 239         skb->ip_summed = CHECKSUM_NONE;
 240
 241         /* If there is no outer header we can fake a checksum offload
 242          * due to the fact that we have already done the checksum in
 243          * software prior to segmenting the frame.
 244          */
 245         if (!skb->encap_hdr_csum)
 246                 features |= NETIF_F_HW_CSUM;
 247
 248         /* Fragment the skb. IP headers of the fragments are updated in
 249          * inet_gso_segment()
 250          */
 251         segs = skb_segment(skb, features);
 252 out:
 253         return segs;
 254 }
 255
 256 int udp_add_offload(struct net *net, struct udp_offload *uo)
 257 {
 258         struct udp_offload_priv *new_offload = kzalloc(sizeof(*new_offload), GFP_ATOMIC);
 259
 260         if (!new_offload)
 261                 return -ENOMEM;
 262
 263         write_pnet(&new_offload->net, net);
 264         new_offload->offload = uo;
 265
 266         spin_lock(&udp_offload_lock);
 267         new_offload->next = udp_offload_base;
 268         rcu_assign_pointer(udp_offload_base, new_offload);
 269         spin_unlock(&udp_offload_lock);
 270
 271         return 0;
 272 }
 273 EXPORT_SYMBOL(udp_add_offload);
 274
 275 static void udp_offload_free_routine(struct rcu_head *head)
 276 {
 277         struct udp_offload_priv *ou_priv = container_of(head, struct udp_offload_priv, rcu);
 278         kfree(ou_priv);
 279 }
 280
 281 void udp_del_offload(struct udp_offload *uo)
 282 {
 283         struct udp_offload_priv __rcu **head = &udp_offload_base;
 284         struct udp_offload_priv *uo_priv;
 285
 286         spin_lock(&udp_offload_lock);
 287
 288         uo_priv = udp_deref_protected(*head);
 289         for (; uo_priv != NULL;
 290              uo_priv = udp_deref_protected(*head)) {
 291                 if (uo_priv->offload == uo) {
 292                         rcu_assign_pointer(*head,
 293                                            udp_deref_protected(uo_priv->next));
 294                         goto unlock;
 295                 }
 296                 head = &uo_priv->next;
 297         }
 298         pr_warn("udp_del_offload: didn't find offload for port %d\n", ntohs(uo->port));
 299 unlock:
 300         spin_unlock(&udp_offload_lock);
 301         if (uo_priv)
 302                 call_rcu(&uo_priv->rcu, udp_offload_free_routine);
 303 }
 304 EXPORT_SYMBOL(udp_del_offload);
 305
 306 struct sk_buff **udp_gro_receive(struct sk_buff **head, struct sk_buff *skb,
 307                                  struct udphdr *uh)
 308 {
 309         struct udp_offload_priv *uo_priv;
 310         struct sk_buff *p, **pp = NULL;
 311         struct udphdr *uh2;
 312         unsigned int off = skb_gro_offset(skb);
 313         int flush = 1;
 314
 315         if (NAPI_GRO_CB(skb)->encap_mark ||
 316             (skb->ip_summed != CHECKSUM_PARTIAL &&
 317              NAPI_GRO_CB(skb)->csum_cnt == 0 &&
 318              !NAPI_GRO_CB(skb)->csum_valid))
 319                 goto out;
 320
 321         /* mark that this skb passed once through the tunnel gro layer */
 322         NAPI_GRO_CB(skb)->encap_mark = 1;
 323
 324         rcu_read_lock();
 325         uo_priv = rcu_dereference(udp_offload_base);
 326         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 327                 if (net_eq(read_pnet(&uo_priv->net), dev_net(skb->dev)) &&
 328                     uo_priv->offload->port == uh->dest &&
 329                     uo_priv->offload->callbacks.gro_receive)
 330                         goto unflush;
 331         }
 332         goto out_unlock;
 333
 334 unflush:
 335         flush = 0;
 336
 337         for (p = *head; p; p = p->next) {
 338                 if (!NAPI_GRO_CB(p)->same_flow)
 339                         continue;
 340
 341                 uh2 = (struct udphdr   *)(p->data + off);
 342
 343                 /* Match ports and either checksums are either both zero
 344                  * or nonzero.
 345                  */
 346                 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
 347                     (!uh->check ^ !uh2->check)) {
 348                         NAPI_GRO_CB(p)->same_flow = 0;
 349                         continue;
 350                 }
 351         }
 352
 353         skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
 354         skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
 355         NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
 356         pp = uo_priv->offload->callbacks.gro_receive(head, skb,
 357                                                      uo_priv->offload);
 358
 359 out_unlock:
 360         rcu_read_unlock();
 361 out:
 362         NAPI_GRO_CB(skb)->flush |= flush;
 363         return pp;
 364 }
 365
 366 static struct sk_buff **udp4_gro_receive(struct sk_buff **head,
 367                                          struct sk_buff *skb)
 368 {
 369         struct udphdr *uh = udp_gro_udphdr(skb);
 370
 371         if (unlikely(!uh))
 372                 goto flush;
 373
 374         /* Don't bother verifying checksum if we're going to flush anyway. */
 375         if (NAPI_GRO_CB(skb)->flush)
 376                 goto skip;
 377
 378         if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
 379                                                  inet_gro_compute_pseudo))
 380                 goto flush;
 381         else if (uh->check)
 382                 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, uh->check,
 383                                              inet_gro_compute_pseudo);
 384 skip:
 385         NAPI_GRO_CB(skb)->is_ipv6 = 0;
 386         return udp_gro_receive(head, skb, uh);
 387
 388 flush:
 389         NAPI_GRO_CB(skb)->flush = 1;
 390         return NULL;
 391 }
 392
 393 int udp_gro_complete(struct sk_buff *skb, int nhoff)
 394 {
 395         struct udp_offload_priv *uo_priv;
 396         __be16 newlen = htons(skb->len - nhoff);
 397         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 398         int err = -ENOSYS;
 399
 400         uh->len = newlen;
 401
 402         rcu_read_lock();
 403
 404         uo_priv = rcu_dereference(udp_offload_base);
 405         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 406                 if (net_eq(read_pnet(&uo_priv->net), dev_net(skb->dev)) &&
 407                     uo_priv->offload->port == uh->dest &&
 408                     uo_priv->offload->callbacks.gro_complete)
 409                         break;
 410         }
 411
 412         if (uo_priv) {
 413                 NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
 414                 err = uo_priv->offload->callbacks.gro_complete(skb,
 415                                 nhoff + sizeof(struct udphdr),
 416                                 uo_priv->offload);
 417         }
 418
 419         rcu_read_unlock();
 420
 421         if (skb->remcsum_offload)
 422                 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
 423
 424         skb->encapsulation = 1;
 425         skb_set_inner_mac_header(skb, nhoff + sizeof(struct udphdr));
 426
 427         return err;
 428 }
 429
 430 static int udp4_gro_complete(struct sk_buff *skb, int nhoff)
 431 {
 432         const struct iphdr *iph = ip_hdr(skb);
 433         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 434
 435         if (uh->check) {
 436                 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL_CSUM;
 437                 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
 438                                           iph->daddr, 0);
 439         } else {
 440                 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL;
 441         }
 442
 443         return udp_gro_complete(skb, nhoff);
 444 }
 445
 446 static const struct net_offload udpv4_offload = {
 447         .callbacks = {
 448                 .gso_segment = udp4_ufo_fragment,
 449                 .gro_receive  = udp4_gro_receive,
 450                 .gro_complete = udp4_gro_complete,
 451         },
 452 };
 453
 454 int __init udpv4_offload_init(void)
 455 {
 456         return inet_add_offload(&udpv4_offload, IPPROTO_UDP);
 457 }