kernel/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         struct rcu_head         rcu;
  25         struct udp_offload_priv __rcu *next;
  26 };
  27
  28 static struct sk_buff *__skb_udp_tunnel_segment(struct sk_buff *skb,
  29         netdev_features_t features,
  30         struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
  31                                              netdev_features_t features),
  32         __be16 new_protocol, bool is_ipv6)
  33 {
  34         struct sk_buff *segs = ERR_PTR(-EINVAL);
  35         u16 mac_offset = skb->mac_header;
  36         int mac_len = skb->mac_len;
  37         int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
  38         __be16 protocol = skb->protocol;
  39         netdev_features_t enc_features;
  40         int udp_offset, outer_hlen;
  41         unsigned int oldlen;
  42         bool need_csum = !!(skb_shinfo(skb)->gso_type &
  43                             SKB_GSO_UDP_TUNNEL_CSUM);
  44         bool remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
  45         bool offload_csum = false, dont_encap = (need_csum || remcsum);
  46
  47         oldlen = (u16)~skb->len;
  48
  49         if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
  50                 goto out;
  51
  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         skb->encap_hdr_csum = need_csum;
  59         skb->remcsum_offload = remcsum;
  60
  61         /* Try to offload checksum if possible */
  62         offload_csum = !!(need_csum &&
  63                           (skb->dev->features &
  64                            (is_ipv6 ? NETIF_F_V6_CSUM : NETIF_F_V4_CSUM)));
  65
  66         /* segment inner packet. */
  67         enc_features = skb->dev->hw_enc_features & features;
  68         segs = gso_inner_segment(skb, enc_features);
  69         if (IS_ERR_OR_NULL(segs)) {
  70                 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
  71                                      mac_len);
  72                 goto out;
  73         }
  74
  75         outer_hlen = skb_tnl_header_len(skb);
  76         udp_offset = outer_hlen - tnl_hlen;
  77         skb = segs;
  78         do {
  79                 struct udphdr *uh;
  80                 int len;
  81                 __be32 delta;
  82
  83                 if (dont_encap) {
  84                         skb->encapsulation = 0;
  85                         skb->ip_summed = CHECKSUM_NONE;
  86                 } else {
  87                         /* Only set up inner headers if we might be offloading
  88                          * inner checksum.
  89                          */
  90                         skb_reset_inner_headers(skb);
  91                         skb->encapsulation = 1;
  92                 }
  93
  94                 skb->mac_len = mac_len;
  95                 skb->protocol = protocol;
  96
  97                 skb_push(skb, outer_hlen);
  98                 skb_reset_mac_header(skb);
  99                 skb_set_network_header(skb, mac_len);
 100                 skb_set_transport_header(skb, udp_offset);
 101                 len = skb->len - udp_offset;
 102                 uh = udp_hdr(skb);
 103                 uh->len = htons(len);
 104
 105                 if (!need_csum)
 106                         continue;
 107
 108                 delta = htonl(oldlen + len);
 109
 110                 uh->check = ~csum_fold((__force __wsum)
 111                                        ((__force u32)uh->check +
 112                                         (__force u32)delta));
 113                 if (offload_csum) {
 114                         skb->ip_summed = CHECKSUM_PARTIAL;
 115                         skb->csum_start = skb_transport_header(skb) - skb->head;
 116                         skb->csum_offset = offsetof(struct udphdr, check);
 117                 } else if (remcsum) {
 118                         /* Need to calculate checksum from scratch,
 119                          * inner checksums are never when doing
 120                          * remote_checksum_offload.
 121                          */
 122
 123                         skb->csum = skb_checksum(skb, udp_offset,
 124                                                  skb->len - udp_offset,
 125                                                  0);
 126                         uh->check = csum_fold(skb->csum);
 127                         if (uh->check == 0)
 128                                 uh->check = CSUM_MANGLED_0;
 129                 } else {
 130                         uh->check = gso_make_checksum(skb, ~uh->check);
 131
 132                         if (uh->check == 0)
 133                                 uh->check = CSUM_MANGLED_0;
 134                 }
 135         } while ((skb = skb->next));
 136 out:
 137         return segs;
 138 }
 139
 140 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
 141                                        netdev_features_t features,
 142                                        bool is_ipv6)
 143 {
 144         __be16 protocol = skb->protocol;
 145         const struct net_offload **offloads;
 146         const struct net_offload *ops;
 147         struct sk_buff *segs = ERR_PTR(-EINVAL);
 148         struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
 149                                              netdev_features_t features);
 150
 151         rcu_read_lock();
 152
 153         switch (skb->inner_protocol_type) {
 154         case ENCAP_TYPE_ETHER:
 155                 protocol = skb->inner_protocol;
 156                 gso_inner_segment = skb_mac_gso_segment;
 157                 break;
 158         case ENCAP_TYPE_IPPROTO:
 159                 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
 160                 ops = rcu_dereference(offloads[skb->inner_ipproto]);
 161                 if (!ops || !ops->callbacks.gso_segment)
 162                         goto out_unlock;
 163                 gso_inner_segment = ops->callbacks.gso_segment;
 164                 break;
 165         default:
 166                 goto out_unlock;
 167         }
 168
 169         segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
 170                                         protocol, is_ipv6);
 171
 172 out_unlock:
 173         rcu_read_unlock();
 174
 175         return segs;
 176 }
 177
 178 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
 179                                          netdev_features_t features)
 180 {
 181         struct sk_buff *segs = ERR_PTR(-EINVAL);
 182         unsigned int mss;
 183         __wsum csum;
 184         struct udphdr *uh;
 185         struct iphdr *iph;
 186
 187         if (skb->encapsulation &&
 188             (skb_shinfo(skb)->gso_type &
 189              (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
 190                 segs = skb_udp_tunnel_segment(skb, features, false);
 191                 goto out;
 192         }
 193
 194         if (!pskb_may_pull(skb, sizeof(struct udphdr)))
 195                 goto out;
 196
 197         mss = skb_shinfo(skb)->gso_size;
 198         if (unlikely(skb->len <= mss))
 199                 goto out;
 200
 201         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
 202                 /* Packet is from an untrusted source, reset gso_segs. */
 203                 int type = skb_shinfo(skb)->gso_type;
 204
 205                 if (unlikely(type & ~(SKB_GSO_UDP | SKB_GSO_DODGY |
 206                                       SKB_GSO_UDP_TUNNEL |
 207                                       SKB_GSO_UDP_TUNNEL_CSUM |
 208                                       SKB_GSO_TUNNEL_REMCSUM |
 209                                       SKB_GSO_IPIP |
 210                                       SKB_GSO_GRE | SKB_GSO_GRE_CSUM) ||
 211                              !(type & (SKB_GSO_UDP))))
 212                         goto out;
 213
 214                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
 215
 216                 segs = NULL;
 217                 goto out;
 218         }
 219
 220         /* Do software UFO. Complete and fill in the UDP checksum as
 221          * HW cannot do checksum of UDP packets sent as multiple
 222          * IP fragments.
 223          */
 224
 225         uh = udp_hdr(skb);
 226         iph = ip_hdr(skb);
 227
 228         uh->check = 0;
 229         csum = skb_checksum(skb, 0, skb->len, 0);
 230         uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
 231         if (uh->check == 0)
 232                 uh->check = CSUM_MANGLED_0;
 233
 234         skb->ip_summed = CHECKSUM_NONE;
 235
 236         /* Fragment the skb. IP headers of the fragments are updated in
 237          * inet_gso_segment()
 238          */
 239         segs = skb_segment(skb, features);
 240 out:
 241         return segs;
 242 }
 243
 244 int udp_add_offload(struct udp_offload *uo)
 245 {
 246         struct udp_offload_priv *new_offload = kzalloc(sizeof(*new_offload), GFP_ATOMIC);
 247
 248         if (!new_offload)
 249                 return -ENOMEM;
 250
 251         new_offload->offload = uo;
 252
 253         spin_lock(&udp_offload_lock);
 254         new_offload->next = udp_offload_base;
 255         rcu_assign_pointer(udp_offload_base, new_offload);
 256         spin_unlock(&udp_offload_lock);
 257
 258         return 0;
 259 }
 260 EXPORT_SYMBOL(udp_add_offload);
 261
 262 static void udp_offload_free_routine(struct rcu_head *head)
 263 {
 264         struct udp_offload_priv *ou_priv = container_of(head, struct udp_offload_priv, rcu);
 265         kfree(ou_priv);
 266 }
 267
 268 void udp_del_offload(struct udp_offload *uo)
 269 {
 270         struct udp_offload_priv __rcu **head = &udp_offload_base;
 271         struct udp_offload_priv *uo_priv;
 272
 273         spin_lock(&udp_offload_lock);
 274
 275         uo_priv = udp_deref_protected(*head);
 276         for (; uo_priv != NULL;
 277              uo_priv = udp_deref_protected(*head)) {
 278                 if (uo_priv->offload == uo) {
 279                         rcu_assign_pointer(*head,
 280                                            udp_deref_protected(uo_priv->next));
 281                         goto unlock;
 282                 }
 283                 head = &uo_priv->next;
 284         }
 285         pr_warn("udp_del_offload: didn't find offload for port %d\n", ntohs(uo->port));
 286 unlock:
 287         spin_unlock(&udp_offload_lock);
 288         if (uo_priv)
 289                 call_rcu(&uo_priv->rcu, udp_offload_free_routine);
 290 }
 291 EXPORT_SYMBOL(udp_del_offload);
 292
 293 struct sk_buff **udp_gro_receive(struct sk_buff **head, struct sk_buff *skb,
 294                                  struct udphdr *uh)
 295 {
 296         struct udp_offload_priv *uo_priv;
 297         struct sk_buff *p, **pp = NULL;
 298         struct udphdr *uh2;
 299         unsigned int off = skb_gro_offset(skb);
 300         int flush = 1;
 301
 302         if (NAPI_GRO_CB(skb)->udp_mark ||
 303             (skb->ip_summed != CHECKSUM_PARTIAL &&
 304              NAPI_GRO_CB(skb)->csum_cnt == 0 &&
 305              !NAPI_GRO_CB(skb)->csum_valid))
 306                 goto out;
 307
 308         /* mark that this skb passed once through the udp gro layer */
 309         NAPI_GRO_CB(skb)->udp_mark = 1;
 310
 311         rcu_read_lock();
 312         uo_priv = rcu_dereference(udp_offload_base);
 313         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 314                 if (uo_priv->offload->port == uh->dest &&
 315                     uo_priv->offload->callbacks.gro_receive)
 316                         goto unflush;
 317         }
 318         goto out_unlock;
 319
 320 unflush:
 321         flush = 0;
 322
 323         for (p = *head; p; p = p->next) {
 324                 if (!NAPI_GRO_CB(p)->same_flow)
 325                         continue;
 326
 327                 uh2 = (struct udphdr   *)(p->data + off);
 328
 329                 /* Match ports and either checksums are either both zero
 330                  * or nonzero.
 331                  */
 332                 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
 333                     (!uh->check ^ !uh2->check)) {
 334                         NAPI_GRO_CB(p)->same_flow = 0;
 335                         continue;
 336                 }
 337         }
 338
 339         skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
 340         skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
 341         NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
 342         pp = uo_priv->offload->callbacks.gro_receive(head, skb,
 343                                                      uo_priv->offload);
 344
 345 out_unlock:
 346         rcu_read_unlock();
 347 out:
 348         NAPI_GRO_CB(skb)->flush |= flush;
 349         return pp;
 350 }
 351
 352 static struct sk_buff **udp4_gro_receive(struct sk_buff **head,
 353                                          struct sk_buff *skb)
 354 {
 355         struct udphdr *uh = udp_gro_udphdr(skb);
 356
 357         if (unlikely(!uh))
 358                 goto flush;
 359
 360         /* Don't bother verifying checksum if we're going to flush anyway. */
 361         if (NAPI_GRO_CB(skb)->flush)
 362                 goto skip;
 363
 364         if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
 365                                                  inet_gro_compute_pseudo))
 366                 goto flush;
 367         else if (uh->check)
 368                 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, uh->check,
 369                                              inet_gro_compute_pseudo);
 370 skip:
 371         NAPI_GRO_CB(skb)->is_ipv6 = 0;
 372         return udp_gro_receive(head, skb, uh);
 373
 374 flush:
 375         NAPI_GRO_CB(skb)->flush = 1;
 376         return NULL;
 377 }
 378
 379 int udp_gro_complete(struct sk_buff *skb, int nhoff)
 380 {
 381         struct udp_offload_priv *uo_priv;
 382         __be16 newlen = htons(skb->len - nhoff);
 383         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 384         int err = -ENOSYS;
 385
 386         uh->len = newlen;
 387
 388         rcu_read_lock();
 389
 390         uo_priv = rcu_dereference(udp_offload_base);
 391         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 392                 if (uo_priv->offload->port == uh->dest &&
 393                     uo_priv->offload->callbacks.gro_complete)
 394                         break;
 395         }
 396
 397         if (uo_priv) {
 398                 NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
 399                 err = uo_priv->offload->callbacks.gro_complete(skb,
 400                                 nhoff + sizeof(struct udphdr),
 401                                 uo_priv->offload);
 402         }
 403
 404         rcu_read_unlock();
 405
 406         if (skb->remcsum_offload)
 407                 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
 408
 409         skb->encapsulation = 1;
 410         skb_set_inner_mac_header(skb, nhoff + sizeof(struct udphdr));
 411
 412         return err;
 413 }
 414
 415 static int udp4_gro_complete(struct sk_buff *skb, int nhoff)
 416 {
 417         const struct iphdr *iph = ip_hdr(skb);
 418         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 419
 420         if (uh->check) {
 421                 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL_CSUM;
 422                 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
 423                                           iph->daddr, 0);
 424         } else {
 425                 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL;
 426         }
 427
 428         return udp_gro_complete(skb, nhoff);
 429 }
 430
 431 static const struct net_offload udpv4_offload = {
 432         .callbacks = {
 433                 .gso_segment = udp4_ufo_fragment,
 434                 .gro_receive  = udp4_gro_receive,
 435                 .gro_complete = udp4_gro_complete,
 436         },
 437 };
 438
 439 int __init udpv4_offload_init(void)
 440 {
 441         return inet_add_offload(&udpv4_offload, IPPROTO_UDP);
 442 }