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         struct sk_buff *segs = ERR_PTR(-EINVAL);
  36         u16 mac_offset = skb->mac_header;
  37         int mac_len = skb->mac_len;
  38         int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
  39         __be16 protocol = skb->protocol;
  40         netdev_features_t enc_features;
  41         int udp_offset, outer_hlen;
  42         unsigned int oldlen;
  43         bool need_csum = !!(skb_shinfo(skb)->gso_type &
  44                             SKB_GSO_UDP_TUNNEL_CSUM);
  45         bool remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
  46         bool offload_csum = false, dont_encap = (need_csum || remcsum);
  47
  48         oldlen = (u16)~skb->len;
  49
  50         if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
  51                 goto out;
  52
  53         skb->encapsulation = 0;
  54         __skb_pull(skb, tnl_hlen);
  55         skb_reset_mac_header(skb);
  56         skb_set_network_header(skb, skb_inner_network_offset(skb));
  57         skb->mac_len = skb_inner_network_offset(skb);
  58         skb->protocol = new_protocol;
  59         skb->encap_hdr_csum = need_csum;
  60         skb->remcsum_offload = remcsum;
  61
  62         /* Try to offload checksum if possible */
  63         offload_csum = !!(need_csum &&
  64                           (skb->dev->features &
  65                            (is_ipv6 ? NETIF_F_V6_CSUM : NETIF_F_V4_CSUM)));
  66
  67         /* segment inner packet. */
  68         enc_features = skb->dev->hw_enc_features & features;
  69         segs = gso_inner_segment(skb, enc_features);
  70         if (IS_ERR_OR_NULL(segs)) {
  71                 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
  72                                      mac_len);
  73                 goto out;
  74         }
  75
  76         outer_hlen = skb_tnl_header_len(skb);
  77         udp_offset = outer_hlen - tnl_hlen;
  78         skb = segs;
  79         do {
  80                 struct udphdr *uh;
  81                 int len;
  82                 __be32 delta;
  83
  84                 if (dont_encap) {
  85                         skb->encapsulation = 0;
  86                         skb->ip_summed = CHECKSUM_NONE;
  87                 } else {
  88                         /* Only set up inner headers if we might be offloading
  89                          * inner checksum.
  90                          */
  91                         skb_reset_inner_headers(skb);
  92                         skb->encapsulation = 1;
  93                 }
  94
  95                 skb->mac_len = mac_len;
  96                 skb->protocol = protocol;
  97
  98                 skb_push(skb, outer_hlen);
  99                 skb_reset_mac_header(skb);
 100                 skb_set_network_header(skb, mac_len);
 101                 skb_set_transport_header(skb, udp_offset);
 102                 len = skb->len - udp_offset;
 103                 uh = udp_hdr(skb);
 104                 uh->len = htons(len);
 105
 106                 if (!need_csum)
 107                         continue;
 108
 109                 delta = htonl(oldlen + len);
 110
 111                 uh->check = ~csum_fold((__force __wsum)
 112                                        ((__force u32)uh->check +
 113                                         (__force u32)delta));
 114                 if (offload_csum) {
 115                         skb->ip_summed = CHECKSUM_PARTIAL;
 116                         skb->csum_start = skb_transport_header(skb) - skb->head;
 117                         skb->csum_offset = offsetof(struct udphdr, check);
 118                 } else if (remcsum) {
 119                         /* Need to calculate checksum from scratch,
 120                          * inner checksums are never when doing
 121                          * remote_checksum_offload.
 122                          */
 123
 124                         skb->csum = skb_checksum(skb, udp_offset,
 125                                                  skb->len - udp_offset,
 126                                                  0);
 127                         uh->check = csum_fold(skb->csum);
 128                         if (uh->check == 0)
 129                                 uh->check = CSUM_MANGLED_0;
 130                 } else {
 131                         uh->check = gso_make_checksum(skb, ~uh->check);
 132
 133                         if (uh->check == 0)
 134                                 uh->check = CSUM_MANGLED_0;
 135                 }
 136         } while ((skb = skb->next));
 137 out:
 138         return segs;
 139 }
 140
 141 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
 142                                        netdev_features_t features,
 143                                        bool is_ipv6)
 144 {
 145         __be16 protocol = skb->protocol;
 146         const struct net_offload **offloads;
 147         const struct net_offload *ops;
 148         struct sk_buff *segs = ERR_PTR(-EINVAL);
 149         struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
 150                                              netdev_features_t features);
 151
 152         rcu_read_lock();
 153
 154         switch (skb->inner_protocol_type) {
 155         case ENCAP_TYPE_ETHER:
 156                 protocol = skb->inner_protocol;
 157                 gso_inner_segment = skb_mac_gso_segment;
 158                 break;
 159         case ENCAP_TYPE_IPPROTO:
 160                 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
 161                 ops = rcu_dereference(offloads[skb->inner_ipproto]);
 162                 if (!ops || !ops->callbacks.gso_segment)
 163                         goto out_unlock;
 164                 gso_inner_segment = ops->callbacks.gso_segment;
 165                 break;
 166         default:
 167                 goto out_unlock;
 168         }
 169
 170         segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
 171                                         protocol, is_ipv6);
 172
 173 out_unlock:
 174         rcu_read_unlock();
 175
 176         return segs;
 177 }
 178
 179 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
 180                                          netdev_features_t features)
 181 {
 182         struct sk_buff *segs = ERR_PTR(-EINVAL);
 183         unsigned int mss;
 184         __wsum csum;
 185         struct udphdr *uh;
 186         struct iphdr *iph;
 187
 188         if (skb->encapsulation &&
 189             (skb_shinfo(skb)->gso_type &
 190              (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
 191                 segs = skb_udp_tunnel_segment(skb, features, false);
 192                 goto out;
 193         }
 194
 195         if (!pskb_may_pull(skb, sizeof(struct udphdr)))
 196                 goto out;
 197
 198         mss = skb_shinfo(skb)->gso_size;
 199         if (unlikely(skb->len <= mss))
 200                 goto out;
 201
 202         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
 203                 /* Packet is from an untrusted source, reset gso_segs. */
 204                 int type = skb_shinfo(skb)->gso_type;
 205
 206                 if (unlikely(type & ~(SKB_GSO_UDP | SKB_GSO_DODGY |
 207                                       SKB_GSO_UDP_TUNNEL |
 208                                       SKB_GSO_UDP_TUNNEL_CSUM |
 209                                       SKB_GSO_TUNNEL_REMCSUM |
 210                                       SKB_GSO_IPIP |
 211                                       SKB_GSO_GRE | SKB_GSO_GRE_CSUM) ||
 212                              !(type & (SKB_GSO_UDP))))
 213                         goto out;
 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         /* Fragment the skb. IP headers of the fragments are updated in
 238          * inet_gso_segment()
 239          */
 240         segs = skb_segment(skb, features);
 241 out:
 242         return segs;
 243 }
 244
 245 int udp_add_offload(struct net *net, struct udp_offload *uo)
 246 {
 247         struct udp_offload_priv *new_offload = kzalloc(sizeof(*new_offload), GFP_ATOMIC);
 248
 249         if (!new_offload)
 250                 return -ENOMEM;
 251
 252         write_pnet(&new_offload->net, net);
 253         new_offload->offload = uo;
 254
 255         spin_lock(&udp_offload_lock);
 256         new_offload->next = udp_offload_base;
 257         rcu_assign_pointer(udp_offload_base, new_offload);
 258         spin_unlock(&udp_offload_lock);
 259
 260         return 0;
 261 }
 262 EXPORT_SYMBOL(udp_add_offload);
 263
 264 static void udp_offload_free_routine(struct rcu_head *head)
 265 {
 266         struct udp_offload_priv *ou_priv = container_of(head, struct udp_offload_priv, rcu);
 267         kfree(ou_priv);
 268 }
 269
 270 void udp_del_offload(struct udp_offload *uo)
 271 {
 272         struct udp_offload_priv __rcu **head = &udp_offload_base;
 273         struct udp_offload_priv *uo_priv;
 274
 275         spin_lock(&udp_offload_lock);
 276
 277         uo_priv = udp_deref_protected(*head);
 278         for (; uo_priv != NULL;
 279              uo_priv = udp_deref_protected(*head)) {
 280                 if (uo_priv->offload == uo) {
 281                         rcu_assign_pointer(*head,
 282                                            udp_deref_protected(uo_priv->next));
 283                         goto unlock;
 284                 }
 285                 head = &uo_priv->next;
 286         }
 287         pr_warn("udp_del_offload: didn't find offload for port %d\n", ntohs(uo->port));
 288 unlock:
 289         spin_unlock(&udp_offload_lock);
 290         if (uo_priv)
 291                 call_rcu(&uo_priv->rcu, udp_offload_free_routine);
 292 }
 293 EXPORT_SYMBOL(udp_del_offload);
 294
 295 struct sk_buff **udp_gro_receive(struct sk_buff **head, struct sk_buff *skb,
 296                                  struct udphdr *uh)
 297 {
 298         struct udp_offload_priv *uo_priv;
 299         struct sk_buff *p, **pp = NULL;
 300         struct udphdr *uh2;
 301         unsigned int off = skb_gro_offset(skb);
 302         int flush = 1;
 303
 304         if (NAPI_GRO_CB(skb)->udp_mark ||
 305             (skb->ip_summed != CHECKSUM_PARTIAL &&
 306              NAPI_GRO_CB(skb)->csum_cnt == 0 &&
 307              !NAPI_GRO_CB(skb)->csum_valid))
 308                 goto out;
 309
 310         /* mark that this skb passed once through the udp gro layer */
 311         NAPI_GRO_CB(skb)->udp_mark = 1;
 312
 313         rcu_read_lock();
 314         uo_priv = rcu_dereference(udp_offload_base);
 315         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 316                 if (net_eq(read_pnet(&uo_priv->net), dev_net(skb->dev)) &&
 317                     uo_priv->offload->port == uh->dest &&
 318                     uo_priv->offload->callbacks.gro_receive)
 319                         goto unflush;
 320         }
 321         goto out_unlock;
 322
 323 unflush:
 324         flush = 0;
 325
 326         for (p = *head; p; p = p->next) {
 327                 if (!NAPI_GRO_CB(p)->same_flow)
 328                         continue;
 329
 330                 uh2 = (struct udphdr   *)(p->data + off);
 331
 332                 /* Match ports and either checksums are either both zero
 333                  * or nonzero.
 334                  */
 335                 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
 336                     (!uh->check ^ !uh2->check)) {
 337                         NAPI_GRO_CB(p)->same_flow = 0;
 338                         continue;
 339                 }
 340         }
 341
 342         skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
 343         skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
 344         NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
 345         pp = uo_priv->offload->callbacks.gro_receive(head, skb,
 346                                                      uo_priv->offload);
 347
 348 out_unlock:
 349         rcu_read_unlock();
 350 out:
 351         NAPI_GRO_CB(skb)->flush |= flush;
 352         return pp;
 353 }
 354
 355 static struct sk_buff **udp4_gro_receive(struct sk_buff **head,
 356                                          struct sk_buff *skb)
 357 {
 358         struct udphdr *uh = udp_gro_udphdr(skb);
 359
 360         if (unlikely(!uh))
 361                 goto flush;
 362
 363         /* Don't bother verifying checksum if we're going to flush anyway. */
 364         if (NAPI_GRO_CB(skb)->flush)
 365                 goto skip;
 366
 367         if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
 368                                                  inet_gro_compute_pseudo))
 369                 goto flush;
 370         else if (uh->check)
 371                 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, uh->check,
 372                                              inet_gro_compute_pseudo);
 373 skip:
 374         NAPI_GRO_CB(skb)->is_ipv6 = 0;
 375         return udp_gro_receive(head, skb, uh);
 376
 377 flush:
 378         NAPI_GRO_CB(skb)->flush = 1;
 379         return NULL;
 380 }
 381
 382 int udp_gro_complete(struct sk_buff *skb, int nhoff)
 383 {
 384         struct udp_offload_priv *uo_priv;
 385         __be16 newlen = htons(skb->len - nhoff);
 386         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 387         int err = -ENOSYS;
 388
 389         uh->len = newlen;
 390
 391         rcu_read_lock();
 392
 393         uo_priv = rcu_dereference(udp_offload_base);
 394         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 395                 if (net_eq(read_pnet(&uo_priv->net), dev_net(skb->dev)) &&
 396                     uo_priv->offload->port == uh->dest &&
 397                     uo_priv->offload->callbacks.gro_complete)
 398                         break;
 399         }
 400
 401         if (uo_priv) {
 402                 NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
 403                 err = uo_priv->offload->callbacks.gro_complete(skb,
 404                                 nhoff + sizeof(struct udphdr),
 405                                 uo_priv->offload);
 406         }
 407
 408         rcu_read_unlock();
 409
 410         if (skb->remcsum_offload)
 411                 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
 412
 413         skb->encapsulation = 1;
 414         skb_set_inner_mac_header(skb, nhoff + sizeof(struct udphdr));
 415
 416         return err;
 417 }
 418
 419 static int udp4_gro_complete(struct sk_buff *skb, int nhoff)
 420 {
 421         const struct iphdr *iph = ip_hdr(skb);
 422         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 423
 424         if (uh->check) {
 425                 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL_CSUM;
 426                 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
 427                                           iph->daddr, 0);
 428         } else {
 429                 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL;
 430         }
 431
 432         return udp_gro_complete(skb, nhoff);
 433 }
 434
 435 static const struct net_offload udpv4_offload = {
 436         .callbacks = {
 437                 .gso_segment = udp4_ufo_fragment,
 438                 .gro_receive  = udp4_gro_receive,
 439                 .gro_complete = udp4_gro_complete,
 440         },
 441 };
 442
 443 int __init udpv4_offload_init(void)
 444 {
 445         return inet_add_offload(&udpv4_offload, IPPROTO_UDP);
 446 }