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