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)
  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;
  43
  44         oldlen = (u16)~skb->len;
  45
  46         if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
  47                 goto out;
  48
  49         skb->encapsulation = 0;
  50         __skb_pull(skb, tnl_hlen);
  51         skb_reset_mac_header(skb);
  52         skb_set_network_header(skb, skb_inner_network_offset(skb));
  53         skb->mac_len = skb_inner_network_offset(skb);
  54         skb->protocol = new_protocol;
  55
  56         need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
  57         if (need_csum)
  58                 skb->encap_hdr_csum = 1;
  59
  60         /* segment inner packet. */
  61         enc_features = skb->dev->hw_enc_features & features;
  62         segs = gso_inner_segment(skb, enc_features);
  63         if (IS_ERR_OR_NULL(segs)) {
  64                 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
  65                                      mac_len);
  66                 goto out;
  67         }
  68
  69         outer_hlen = skb_tnl_header_len(skb);
  70         udp_offset = outer_hlen - tnl_hlen;
  71         skb = segs;
  72         do {
  73                 struct udphdr *uh;
  74                 int len;
  75
  76                 skb_reset_inner_headers(skb);
  77                 skb->encapsulation = 1;
  78
  79                 skb->mac_len = mac_len;
  80
  81                 skb_push(skb, outer_hlen);
  82                 skb_reset_mac_header(skb);
  83                 skb_set_network_header(skb, mac_len);
  84                 skb_set_transport_header(skb, udp_offset);
  85                 len = skb->len - udp_offset;
  86                 uh = udp_hdr(skb);
  87                 uh->len = htons(len);
  88
  89                 if (need_csum) {
  90                         __be32 delta = htonl(oldlen + len);
  91
  92                         uh->check = ~csum_fold((__force __wsum)
  93                                                ((__force u32)uh->check +
  94                                                 (__force u32)delta));
  95                         uh->check = gso_make_checksum(skb, ~uh->check);
  96
  97                         if (uh->check == 0)
  98                                 uh->check = CSUM_MANGLED_0;
  99                 }
 100
 101                 skb->protocol = protocol;
 102         } while ((skb = skb->next));
 103 out:
 104         return segs;
 105 }
 106
 107 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
 108                                        netdev_features_t features,
 109                                        bool is_ipv6)
 110 {
 111         __be16 protocol = skb->protocol;
 112         const struct net_offload **offloads;
 113         const struct net_offload *ops;
 114         struct sk_buff *segs = ERR_PTR(-EINVAL);
 115         struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
 116                                              netdev_features_t features);
 117
 118         rcu_read_lock();
 119
 120         switch (skb->inner_protocol_type) {
 121         case ENCAP_TYPE_ETHER:
 122                 protocol = skb->inner_protocol;
 123                 gso_inner_segment = skb_mac_gso_segment;
 124                 break;
 125         case ENCAP_TYPE_IPPROTO:
 126                 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
 127                 ops = rcu_dereference(offloads[skb->inner_ipproto]);
 128                 if (!ops || !ops->callbacks.gso_segment)
 129                         goto out_unlock;
 130                 gso_inner_segment = ops->callbacks.gso_segment;
 131                 break;
 132         default:
 133                 goto out_unlock;
 134         }
 135
 136         segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
 137                                         protocol);
 138
 139 out_unlock:
 140         rcu_read_unlock();
 141
 142         return segs;
 143 }
 144
 145 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
 146                                          netdev_features_t features)
 147 {
 148         struct sk_buff *segs = ERR_PTR(-EINVAL);
 149         unsigned int mss;
 150         __wsum csum;
 151         struct udphdr *uh;
 152         struct iphdr *iph;
 153
 154         if (skb->encapsulation &&
 155             (skb_shinfo(skb)->gso_type &
 156              (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
 157                 segs = skb_udp_tunnel_segment(skb, features, false);
 158                 goto out;
 159         }
 160
 161         if (!pskb_may_pull(skb, sizeof(struct udphdr)))
 162                 goto out;
 163
 164         mss = skb_shinfo(skb)->gso_size;
 165         if (unlikely(skb->len <= mss))
 166                 goto out;
 167
 168         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
 169                 /* Packet is from an untrusted source, reset gso_segs. */
 170                 int type = skb_shinfo(skb)->gso_type;
 171
 172                 if (unlikely(type & ~(SKB_GSO_UDP | SKB_GSO_DODGY |
 173                                       SKB_GSO_UDP_TUNNEL |
 174                                       SKB_GSO_UDP_TUNNEL_CSUM |
 175                                       SKB_GSO_IPIP |
 176                                       SKB_GSO_GRE | SKB_GSO_GRE_CSUM |
 177                                       SKB_GSO_MPLS) ||
 178                              !(type & (SKB_GSO_UDP))))
 179                         goto out;
 180
 181                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
 182
 183                 segs = NULL;
 184                 goto out;
 185         }
 186
 187         /* Do software UFO. Complete and fill in the UDP checksum as
 188          * HW cannot do checksum of UDP packets sent as multiple
 189          * IP fragments.
 190          */
 191
 192         uh = udp_hdr(skb);
 193         iph = ip_hdr(skb);
 194
 195         uh->check = 0;
 196         csum = skb_checksum(skb, 0, skb->len, 0);
 197         uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
 198         if (uh->check == 0)
 199                 uh->check = CSUM_MANGLED_0;
 200
 201         skb->ip_summed = CHECKSUM_NONE;
 202
 203         /* Fragment the skb. IP headers of the fragments are updated in
 204          * inet_gso_segment()
 205          */
 206         segs = skb_segment(skb, features);
 207 out:
 208         return segs;
 209 }
 210
 211 int udp_add_offload(struct udp_offload *uo)
 212 {
 213         struct udp_offload_priv *new_offload = kzalloc(sizeof(*new_offload), GFP_ATOMIC);
 214
 215         if (!new_offload)
 216                 return -ENOMEM;
 217
 218         new_offload->offload = uo;
 219
 220         spin_lock(&udp_offload_lock);
 221         new_offload->next = udp_offload_base;
 222         rcu_assign_pointer(udp_offload_base, new_offload);
 223         spin_unlock(&udp_offload_lock);
 224
 225         return 0;
 226 }
 227 EXPORT_SYMBOL(udp_add_offload);
 228
 229 static void udp_offload_free_routine(struct rcu_head *head)
 230 {
 231         struct udp_offload_priv *ou_priv = container_of(head, struct udp_offload_priv, rcu);
 232         kfree(ou_priv);
 233 }
 234
 235 void udp_del_offload(struct udp_offload *uo)
 236 {
 237         struct udp_offload_priv __rcu **head = &udp_offload_base;
 238         struct udp_offload_priv *uo_priv;
 239
 240         spin_lock(&udp_offload_lock);
 241
 242         uo_priv = udp_deref_protected(*head);
 243         for (; uo_priv != NULL;
 244              uo_priv = udp_deref_protected(*head)) {
 245                 if (uo_priv->offload == uo) {
 246                         rcu_assign_pointer(*head,
 247                                            udp_deref_protected(uo_priv->next));
 248                         goto unlock;
 249                 }
 250                 head = &uo_priv->next;
 251         }
 252         pr_warn("udp_del_offload: didn't find offload for port %d\n", ntohs(uo->port));
 253 unlock:
 254         spin_unlock(&udp_offload_lock);
 255         if (uo_priv != NULL)
 256                 call_rcu(&uo_priv->rcu, udp_offload_free_routine);
 257 }
 258 EXPORT_SYMBOL(udp_del_offload);
 259
 260 struct sk_buff **udp_gro_receive(struct sk_buff **head, struct sk_buff *skb,
 261                                  struct udphdr *uh)
 262 {
 263         struct udp_offload_priv *uo_priv;
 264         struct sk_buff *p, **pp = NULL;
 265         struct udphdr *uh2;
 266         unsigned int off = skb_gro_offset(skb);
 267         int flush = 1;
 268
 269         if (NAPI_GRO_CB(skb)->udp_mark ||
 270             (skb->ip_summed != CHECKSUM_PARTIAL &&
 271              NAPI_GRO_CB(skb)->csum_cnt == 0 &&
 272              !NAPI_GRO_CB(skb)->csum_valid))
 273                 goto out;
 274
 275         /* mark that this skb passed once through the udp gro layer */
 276         NAPI_GRO_CB(skb)->udp_mark = 1;
 277
 278         rcu_read_lock();
 279         uo_priv = rcu_dereference(udp_offload_base);
 280         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 281                 if (uo_priv->offload->port == uh->dest &&
 282                     uo_priv->offload->callbacks.gro_receive)
 283                         goto unflush;
 284         }
 285         goto out_unlock;
 286
 287 unflush:
 288         flush = 0;
 289
 290         for (p = *head; p; p = p->next) {
 291                 if (!NAPI_GRO_CB(p)->same_flow)
 292                         continue;
 293
 294                 uh2 = (struct udphdr   *)(p->data + off);
 295
 296                 /* Match ports and either checksums are either both zero
 297                  * or nonzero.
 298                  */
 299                 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
 300                     (!uh->check ^ !uh2->check)) {
 301                         NAPI_GRO_CB(p)->same_flow = 0;
 302                         continue;
 303                 }
 304         }
 305
 306         skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
 307         skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
 308         NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
 309         pp = uo_priv->offload->callbacks.gro_receive(head, skb);
 310
 311 out_unlock:
 312         rcu_read_unlock();
 313 out:
 314         NAPI_GRO_CB(skb)->flush |= flush;
 315         return pp;
 316 }
 317
 318 static struct sk_buff **udp4_gro_receive(struct sk_buff **head,
 319                                          struct sk_buff *skb)
 320 {
 321         struct udphdr *uh = udp_gro_udphdr(skb);
 322
 323         if (unlikely(!uh))
 324                 goto flush;
 325
 326         /* Don't bother verifying checksum if we're going to flush anyway. */
 327         if (NAPI_GRO_CB(skb)->flush)
 328                 goto skip;
 329
 330         if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
 331                                                  inet_gro_compute_pseudo))
 332                 goto flush;
 333         else if (uh->check)
 334                 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, uh->check,
 335                                              inet_gro_compute_pseudo);
 336 skip:
 337         NAPI_GRO_CB(skb)->is_ipv6 = 0;
 338         return udp_gro_receive(head, skb, uh);
 339
 340 flush:
 341         NAPI_GRO_CB(skb)->flush = 1;
 342         return NULL;
 343 }
 344
 345 int udp_gro_complete(struct sk_buff *skb, int nhoff)
 346 {
 347         struct udp_offload_priv *uo_priv;
 348         __be16 newlen = htons(skb->len - nhoff);
 349         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 350         int err = -ENOSYS;
 351
 352         uh->len = newlen;
 353
 354         rcu_read_lock();
 355
 356         uo_priv = rcu_dereference(udp_offload_base);
 357         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 358                 if (uo_priv->offload->port == uh->dest &&
 359                     uo_priv->offload->callbacks.gro_complete)
 360                         break;
 361         }
 362
 363         if (uo_priv != NULL) {
 364                 NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
 365                 err = uo_priv->offload->callbacks.gro_complete(skb, nhoff + sizeof(struct udphdr));
 366         }
 367
 368         rcu_read_unlock();
 369         return err;
 370 }
 371
 372 static int udp4_gro_complete(struct sk_buff *skb, int nhoff)
 373 {
 374         const struct iphdr *iph = ip_hdr(skb);
 375         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 376
 377         if (uh->check)
 378                 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
 379                                           iph->daddr, 0);
 380
 381         return udp_gro_complete(skb, nhoff);
 382 }
 383
 384 static const struct net_offload udpv4_offload = {
 385         .callbacks = {
 386                 .gso_segment = udp4_ufo_fragment,
 387                 .gro_receive  = udp4_gro_receive,
 388                 .gro_complete = udp4_gro_complete,
 389         },
 390 };
 391
 392 int __init udpv4_offload_init(void)
 393 {
 394         return inet_add_offload(&udpv4_offload, IPPROTO_UDP);
 395 }