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