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 & NETIF_F_HW_CSUM) ||
  65                            (skb->dev->features & (is_ipv6 ?
  66                             NETIF_F_IPV6_CSUM : NETIF_F_IP_CSUM))));
  67
  68         /* segment inner packet. */
  69         enc_features = skb->dev->hw_enc_features & features;
  70         segs = gso_inner_segment(skb, enc_features);
  71         if (IS_ERR_OR_NULL(segs)) {
  72                 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
  73                                      mac_len);
  74                 goto out;
  75         }
  76
  77         outer_hlen = skb_tnl_header_len(skb);
  78         udp_offset = outer_hlen - tnl_hlen;
  79         skb = segs;
  80         do {
  81                 struct udphdr *uh;
  82                 int len;
  83                 __be32 delta;
  84
  85                 if (dont_encap) {
  86                         skb->encapsulation = 0;
  87                         skb->ip_summed = CHECKSUM_NONE;
  88                 } else {
  89                         /* Only set up inner headers if we might be offloading
  90                          * inner checksum.
  91                          */
  92                         skb_reset_inner_headers(skb);
  93                         skb->encapsulation = 1;
  94                 }
  95
  96                 skb->mac_len = mac_len;
  97                 skb->protocol = protocol;
  98
  99                 skb_push(skb, outer_hlen);
 100                 skb_reset_mac_header(skb);
 101                 skb_set_network_header(skb, mac_len);
 102                 skb_set_transport_header(skb, udp_offset);
 103                 len = skb->len - udp_offset;
 104                 uh = udp_hdr(skb);
 105                 uh->len = htons(len);
 106
 107                 if (!need_csum)
 108                         continue;
 109
 110                 delta = htonl(oldlen + len);
 111
 112                 uh->check = ~csum_fold((__force __wsum)
 113                                        ((__force u32)uh->check +
 114                                         (__force u32)delta));
 115                 if (offload_csum) {
 116                         skb->ip_summed = CHECKSUM_PARTIAL;
 117                         skb->csum_start = skb_transport_header(skb) - skb->head;
 118                         skb->csum_offset = offsetof(struct udphdr, check);
 119                 } else if (remcsum) {
 120                         /* Need to calculate checksum from scratch,
 121                          * inner checksums are never when doing
 122                          * remote_checksum_offload.
 123                          */
 124
 125                         skb->csum = skb_checksum(skb, udp_offset,
 126                                                  skb->len - udp_offset,
 127                                                  0);
 128                         uh->check = csum_fold(skb->csum);
 129                         if (uh->check == 0)
 130                                 uh->check = CSUM_MANGLED_0;
 131                 } else {
 132                         uh->check = gso_make_checksum(skb, ~uh->check);
 133
 134                         if (uh->check == 0)
 135                                 uh->check = CSUM_MANGLED_0;
 136                 }
 137         } while ((skb = skb->next));
 138 out:
 139         return segs;
 140 }
 141
 142 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
 143                                        netdev_features_t features,
 144                                        bool is_ipv6)
 145 {
 146         __be16 protocol = skb->protocol;
 147         const struct net_offload **offloads;
 148         const struct net_offload *ops;
 149         struct sk_buff *segs = ERR_PTR(-EINVAL);
 150         struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
 151                                              netdev_features_t features);
 152
 153         rcu_read_lock();
 154
 155         switch (skb->inner_protocol_type) {
 156         case ENCAP_TYPE_ETHER:
 157                 protocol = skb->inner_protocol;
 158                 gso_inner_segment = skb_mac_gso_segment;
 159                 break;
 160         case ENCAP_TYPE_IPPROTO:
 161                 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
 162                 ops = rcu_dereference(offloads[skb->inner_ipproto]);
 163                 if (!ops || !ops->callbacks.gso_segment)
 164                         goto out_unlock;
 165                 gso_inner_segment = ops->callbacks.gso_segment;
 166                 break;
 167         default:
 168                 goto out_unlock;
 169         }
 170
 171         segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
 172                                         protocol, is_ipv6);
 173
 174 out_unlock:
 175         rcu_read_unlock();
 176
 177         return segs;
 178 }
 179
 180 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
 181                                          netdev_features_t features)
 182 {
 183         struct sk_buff *segs = ERR_PTR(-EINVAL);
 184         unsigned int mss;
 185         __wsum csum;
 186         struct udphdr *uh;
 187         struct iphdr *iph;
 188
 189         if (skb->encapsulation &&
 190             (skb_shinfo(skb)->gso_type &
 191              (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
 192                 segs = skb_udp_tunnel_segment(skb, features, false);
 193                 goto out;
 194         }
 195
 196         if (!pskb_may_pull(skb, sizeof(struct udphdr)))
 197                 goto out;
 198
 199         mss = skb_shinfo(skb)->gso_size;
 200         if (unlikely(skb->len <= mss))
 201                 goto out;
 202
 203         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
 204                 /* Packet is from an untrusted source, reset gso_segs. */
 205                 int type = skb_shinfo(skb)->gso_type;
 206
 207                 if (unlikely(type & ~(SKB_GSO_UDP | SKB_GSO_DODGY |
 208                                       SKB_GSO_UDP_TUNNEL |
 209                                       SKB_GSO_UDP_TUNNEL_CSUM |
 210                                       SKB_GSO_TUNNEL_REMCSUM |
 211                                       SKB_GSO_IPIP |
 212                                       SKB_GSO_GRE | SKB_GSO_GRE_CSUM) ||
 213                              !(type & (SKB_GSO_UDP))))
 214                         goto out;
 215
 216                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
 217
 218                 segs = NULL;
 219                 goto out;
 220         }
 221
 222         /* Do software UFO. Complete and fill in the UDP checksum as
 223          * HW cannot do checksum of UDP packets sent as multiple
 224          * IP fragments.
 225          */
 226
 227         uh = udp_hdr(skb);
 228         iph = ip_hdr(skb);
 229
 230         uh->check = 0;
 231         csum = skb_checksum(skb, 0, skb->len, 0);
 232         uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
 233         if (uh->check == 0)
 234                 uh->check = CSUM_MANGLED_0;
 235
 236         skb->ip_summed = CHECKSUM_NONE;
 237
 238         /* Fragment the skb. IP headers of the fragments are updated in
 239          * inet_gso_segment()
 240          */
 241         segs = skb_segment(skb, features);
 242 out:
 243         return segs;
 244 }
 245
 246 int udp_add_offload(struct net *net, struct udp_offload *uo)
 247 {
 248         struct udp_offload_priv *new_offload = kzalloc(sizeof(*new_offload), GFP_ATOMIC);
 249
 250         if (!new_offload)
 251                 return -ENOMEM;
 252
 253         write_pnet(&new_offload->net, net);
 254         new_offload->offload = uo;
 255
 256         spin_lock(&udp_offload_lock);
 257         new_offload->next = udp_offload_base;
 258         rcu_assign_pointer(udp_offload_base, new_offload);
 259         spin_unlock(&udp_offload_lock);
 260
 261         return 0;
 262 }
 263 EXPORT_SYMBOL(udp_add_offload);
 264
 265 static void udp_offload_free_routine(struct rcu_head *head)
 266 {
 267         struct udp_offload_priv *ou_priv = container_of(head, struct udp_offload_priv, rcu);
 268         kfree(ou_priv);
 269 }
 270
 271 void udp_del_offload(struct udp_offload *uo)
 272 {
 273         struct udp_offload_priv __rcu **head = &udp_offload_base;
 274         struct udp_offload_priv *uo_priv;
 275
 276         spin_lock(&udp_offload_lock);
 277
 278         uo_priv = udp_deref_protected(*head);
 279         for (; uo_priv != NULL;
 280              uo_priv = udp_deref_protected(*head)) {
 281                 if (uo_priv->offload == uo) {
 282                         rcu_assign_pointer(*head,
 283                                            udp_deref_protected(uo_priv->next));
 284                         goto unlock;
 285                 }
 286                 head = &uo_priv->next;
 287         }
 288         pr_warn("udp_del_offload: didn't find offload for port %d\n", ntohs(uo->port));
 289 unlock:
 290         spin_unlock(&udp_offload_lock);
 291         if (uo_priv)
 292                 call_rcu(&uo_priv->rcu, udp_offload_free_routine);
 293 }
 294 EXPORT_SYMBOL(udp_del_offload);
 295
 296 struct sk_buff **udp_gro_receive(struct sk_buff **head, struct sk_buff *skb,
 297                                  struct udphdr *uh)
 298 {
 299         struct udp_offload_priv *uo_priv;
 300         struct sk_buff *p, **pp = NULL;
 301         struct udphdr *uh2;
 302         unsigned int off = skb_gro_offset(skb);
 303         int flush = 1;
 304
 305         if (NAPI_GRO_CB(skb)->udp_mark ||
 306             (skb->ip_summed != CHECKSUM_PARTIAL &&
 307              NAPI_GRO_CB(skb)->csum_cnt == 0 &&
 308              !NAPI_GRO_CB(skb)->csum_valid))
 309                 goto out;
 310
 311         /* mark that this skb passed once through the udp gro layer */
 312         NAPI_GRO_CB(skb)->udp_mark = 1;
 313
 314         rcu_read_lock();
 315         uo_priv = rcu_dereference(udp_offload_base);
 316         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 317                 if (net_eq(read_pnet(&uo_priv->net), dev_net(skb->dev)) &&
 318                     uo_priv->offload->port == uh->dest &&
 319                     uo_priv->offload->callbacks.gro_receive)
 320                         goto unflush;
 321         }
 322         goto out_unlock;
 323
 324 unflush:
 325         flush = 0;
 326
 327         for (p = *head; p; p = p->next) {
 328                 if (!NAPI_GRO_CB(p)->same_flow)
 329                         continue;
 330
 331                 uh2 = (struct udphdr   *)(p->data + off);
 332
 333                 /* Match ports and either checksums are either both zero
 334                  * or nonzero.
 335                  */
 336                 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
 337                     (!uh->check ^ !uh2->check)) {
 338                         NAPI_GRO_CB(p)->same_flow = 0;
 339                         continue;
 340                 }
 341         }
 342
 343         skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
 344         skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
 345         NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
 346         pp = uo_priv->offload->callbacks.gro_receive(head, skb,
 347                                                      uo_priv->offload);
 348
 349 out_unlock:
 350         rcu_read_unlock();
 351 out:
 352         NAPI_GRO_CB(skb)->flush |= flush;
 353         return pp;
 354 }
 355
 356 static struct sk_buff **udp4_gro_receive(struct sk_buff **head,
 357                                          struct sk_buff *skb)
 358 {
 359         struct udphdr *uh = udp_gro_udphdr(skb);
 360
 361         if (unlikely(!uh))
 362                 goto flush;
 363
 364         /* Don't bother verifying checksum if we're going to flush anyway. */
 365         if (NAPI_GRO_CB(skb)->flush)
 366                 goto skip;
 367
 368         if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
 369                                                  inet_gro_compute_pseudo))
 370                 goto flush;
 371         else if (uh->check)
 372                 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, uh->check,
 373                                              inet_gro_compute_pseudo);
 374 skip:
 375         NAPI_GRO_CB(skb)->is_ipv6 = 0;
 376         return udp_gro_receive(head, skb, uh);
 377
 378 flush:
 379         NAPI_GRO_CB(skb)->flush = 1;
 380         return NULL;
 381 }
 382
 383 int udp_gro_complete(struct sk_buff *skb, int nhoff)
 384 {
 385         struct udp_offload_priv *uo_priv;
 386         __be16 newlen = htons(skb->len - nhoff);
 387         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 388         int err = -ENOSYS;
 389
 390         uh->len = newlen;
 391
 392         rcu_read_lock();
 393
 394         uo_priv = rcu_dereference(udp_offload_base);
 395         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 396                 if (net_eq(read_pnet(&uo_priv->net), dev_net(skb->dev)) &&
 397                     uo_priv->offload->port == uh->dest &&
 398                     uo_priv->offload->callbacks.gro_complete)
 399                         break;
 400         }
 401
 402         if (uo_priv) {
 403                 NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
 404                 err = uo_priv->offload->callbacks.gro_complete(skb,
 405                                 nhoff + sizeof(struct udphdr),
 406                                 uo_priv->offload);
 407         }
 408
 409         rcu_read_unlock();
 410
 411         if (skb->remcsum_offload)
 412                 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
 413
 414         skb->encapsulation = 1;
 415         skb_set_inner_mac_header(skb, nhoff + sizeof(struct udphdr));
 416
 417         return err;
 418 }
 419
 420 static int udp4_gro_complete(struct sk_buff *skb, int nhoff)
 421 {
 422         const struct iphdr *iph = ip_hdr(skb);
 423         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 424
 425         if (uh->check) {
 426                 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL_CSUM;
 427                 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
 428                                           iph->daddr, 0);
 429         } else {
 430                 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL;
 431         }
 432
 433         return udp_gro_complete(skb, nhoff);
 434 }
 435
 436 static const struct net_offload udpv4_offload = {
 437         .callbacks = {
 438                 .gso_segment = udp4_ufo_fragment,
 439                 .gro_receive  = udp4_gro_receive,
 440                 .gro_complete = udp4_gro_complete,
 441         },
 442 };
 443
 444 int __init udpv4_offload_init(void)
 445 {
 446         return inet_add_offload(&udpv4_offload, IPPROTO_UDP);
 447 }