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