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 int udp4_ufo_send_check(struct sk_buff *skb)
  29 {
  30         return 0;
  31 }
  32
  33 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
  34                                        netdev_features_t features)
  35 {
  36         struct sk_buff *segs = ERR_PTR(-EINVAL);
  37         u16 mac_offset = skb->mac_header;
  38         int mac_len = skb->mac_len;
  39         int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
  40         __be16 protocol = skb->protocol;
  41         netdev_features_t enc_features;
  42         int udp_offset, outer_hlen;
  43         unsigned int oldlen;
  44         bool need_csum;
  45
  46         oldlen = (u16)~skb->len;
  47
  48         if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
  49                 goto out;
  50
  51         skb->encapsulation = 0;
  52         __skb_pull(skb, tnl_hlen);
  53         skb_reset_mac_header(skb);
  54         skb_set_network_header(skb, skb_inner_network_offset(skb));
  55         skb->mac_len = skb_inner_network_offset(skb);
  56         skb->protocol = htons(ETH_P_TEB);
  57
  58         need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
  59         if (need_csum)
  60                 skb->encap_hdr_csum = 1;
  61
  62         /* segment inner packet. */
  63         enc_features = skb->dev->hw_enc_features & netif_skb_features(skb);
  64         segs = skb_mac_gso_segment(skb, enc_features);
  65         if (IS_ERR_OR_NULL(segs)) {
  66                 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
  67                                      mac_len);
  68                 goto out;
  69         }
  70
  71         outer_hlen = skb_tnl_header_len(skb);
  72         udp_offset = outer_hlen - tnl_hlen;
  73         skb = segs;
  74         do {
  75                 struct udphdr *uh;
  76                 int len;
  77
  78                 skb_reset_inner_headers(skb);
  79                 skb->encapsulation = 1;
  80
  81                 skb->mac_len = mac_len;
  82
  83                 skb_push(skb, outer_hlen);
  84                 skb_reset_mac_header(skb);
  85                 skb_set_network_header(skb, mac_len);
  86                 skb_set_transport_header(skb, udp_offset);
  87                 len = skb->len - udp_offset;
  88                 uh = udp_hdr(skb);
  89                 uh->len = htons(len);
  90
  91                 if (need_csum) {
  92                         __be32 delta = htonl(oldlen + len);
  93
  94                         uh->check = ~csum_fold((__force __wsum)
  95                                                ((__force u32)uh->check +
  96                                                 (__force u32)delta));
  97                         uh->check = gso_make_checksum(skb, ~uh->check);
  98
  99                         if (uh->check == 0)
 100                                 uh->check = CSUM_MANGLED_0;
 101                 }
 102
 103                 skb->protocol = protocol;
 104         } while ((skb = skb->next));
 105 out:
 106         return segs;
 107 }
 108
 109 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
 110                                          netdev_features_t features)
 111 {
 112         struct sk_buff *segs = ERR_PTR(-EINVAL);
 113         unsigned int mss;
 114         __wsum csum;
 115         struct udphdr *uh;
 116         struct iphdr *iph;
 117
 118         if (skb->encapsulation &&
 119             (skb_shinfo(skb)->gso_type &
 120              (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
 121                 segs = skb_udp_tunnel_segment(skb, features);
 122                 goto out;
 123         }
 124
 125         if (!pskb_may_pull(skb, sizeof(struct udphdr)))
 126                 goto out;
 127
 128         mss = skb_shinfo(skb)->gso_size;
 129         if (unlikely(skb->len <= mss))
 130                 goto out;
 131
 132         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
 133                 /* Packet is from an untrusted source, reset gso_segs. */
 134                 int type = skb_shinfo(skb)->gso_type;
 135
 136                 if (unlikely(type & ~(SKB_GSO_UDP | SKB_GSO_DODGY |
 137                                       SKB_GSO_UDP_TUNNEL |
 138                                       SKB_GSO_UDP_TUNNEL_CSUM |
 139                                       SKB_GSO_IPIP |
 140                                       SKB_GSO_GRE | SKB_GSO_GRE_CSUM |
 141                                       SKB_GSO_MPLS) ||
 142                              !(type & (SKB_GSO_UDP))))
 143                         goto out;
 144
 145                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
 146
 147                 segs = NULL;
 148                 goto out;
 149         }
 150
 151         /* Do software UFO. Complete and fill in the UDP checksum as
 152          * HW cannot do checksum of UDP packets sent as multiple
 153          * IP fragments.
 154          */
 155
 156         uh = udp_hdr(skb);
 157         iph = ip_hdr(skb);
 158
 159         uh->check = 0;
 160         csum = skb_checksum(skb, 0, skb->len, 0);
 161         uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
 162         if (uh->check == 0)
 163                 uh->check = CSUM_MANGLED_0;
 164
 165         skb->ip_summed = CHECKSUM_NONE;
 166
 167         /* Fragment the skb. IP headers of the fragments are updated in
 168          * inet_gso_segment()
 169          */
 170         segs = skb_segment(skb, features);
 171 out:
 172         return segs;
 173 }
 174
 175 int udp_add_offload(struct udp_offload *uo)
 176 {
 177         struct udp_offload_priv *new_offload = kzalloc(sizeof(*new_offload), GFP_ATOMIC);
 178
 179         if (!new_offload)
 180                 return -ENOMEM;
 181
 182         new_offload->offload = uo;
 183
 184         spin_lock(&udp_offload_lock);
 185         new_offload->next = udp_offload_base;
 186         rcu_assign_pointer(udp_offload_base, new_offload);
 187         spin_unlock(&udp_offload_lock);
 188
 189         return 0;
 190 }
 191 EXPORT_SYMBOL(udp_add_offload);
 192
 193 static void udp_offload_free_routine(struct rcu_head *head)
 194 {
 195         struct udp_offload_priv *ou_priv = container_of(head, struct udp_offload_priv, rcu);
 196         kfree(ou_priv);
 197 }
 198
 199 void udp_del_offload(struct udp_offload *uo)
 200 {
 201         struct udp_offload_priv __rcu **head = &udp_offload_base;
 202         struct udp_offload_priv *uo_priv;
 203
 204         spin_lock(&udp_offload_lock);
 205
 206         uo_priv = udp_deref_protected(*head);
 207         for (; uo_priv != NULL;
 208              uo_priv = udp_deref_protected(*head)) {
 209                 if (uo_priv->offload == uo) {
 210                         rcu_assign_pointer(*head,
 211                                            udp_deref_protected(uo_priv->next));
 212                         goto unlock;
 213                 }
 214                 head = &uo_priv->next;
 215         }
 216         pr_warn("udp_del_offload: didn't find offload for port %d\n", ntohs(uo->port));
 217 unlock:
 218         spin_unlock(&udp_offload_lock);
 219         if (uo_priv != NULL)
 220                 call_rcu(&uo_priv->rcu, udp_offload_free_routine);
 221 }
 222 EXPORT_SYMBOL(udp_del_offload);
 223
 224 struct sk_buff **udp_gro_receive(struct sk_buff **head, struct sk_buff *skb,
 225                                  struct udphdr *uh)
 226 {
 227         struct udp_offload_priv *uo_priv;
 228         struct sk_buff *p, **pp = NULL;
 229         struct udphdr *uh2;
 230         unsigned int off = skb_gro_offset(skb);
 231         int flush = 1;
 232
 233         if (NAPI_GRO_CB(skb)->udp_mark ||
 234             (skb->ip_summed != CHECKSUM_PARTIAL &&
 235              NAPI_GRO_CB(skb)->csum_cnt == 0 &&
 236              !NAPI_GRO_CB(skb)->csum_valid))
 237                 goto out;
 238
 239         /* mark that this skb passed once through the udp gro layer */
 240         NAPI_GRO_CB(skb)->udp_mark = 1;
 241
 242         rcu_read_lock();
 243         uo_priv = rcu_dereference(udp_offload_base);
 244         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 245                 if (uo_priv->offload->port == uh->dest &&
 246                     uo_priv->offload->callbacks.gro_receive)
 247                         goto unflush;
 248         }
 249         goto out_unlock;
 250
 251 unflush:
 252         flush = 0;
 253
 254         for (p = *head; p; p = p->next) {
 255                 if (!NAPI_GRO_CB(p)->same_flow)
 256                         continue;
 257
 258                 uh2 = (struct udphdr   *)(p->data + off);
 259
 260                 /* Match ports and either checksums are either both zero
 261                  * or nonzero.
 262                  */
 263                 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
 264                     (!uh->check ^ !uh2->check)) {
 265                         NAPI_GRO_CB(p)->same_flow = 0;
 266                         continue;
 267                 }
 268         }
 269
 270         skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
 271         skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
 272         NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
 273         pp = uo_priv->offload->callbacks.gro_receive(head, skb);
 274
 275 out_unlock:
 276         rcu_read_unlock();
 277 out:
 278         NAPI_GRO_CB(skb)->flush |= flush;
 279         return pp;
 280 }
 281
 282 static struct sk_buff **udp4_gro_receive(struct sk_buff **head,
 283                                          struct sk_buff *skb)
 284 {
 285         struct udphdr *uh = udp_gro_udphdr(skb);
 286
 287         if (unlikely(!uh))
 288                 goto flush;
 289
 290         /* Don't bother verifying checksum if we're going to flush anyway. */
 291         if (NAPI_GRO_CB(skb)->flush)
 292                 goto skip;
 293
 294         if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
 295                                                  inet_gro_compute_pseudo))
 296                 goto flush;
 297         else if (uh->check)
 298                 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, uh->check,
 299                                              inet_gro_compute_pseudo);
 300 skip:
 301         return udp_gro_receive(head, skb, uh);
 302
 303 flush:
 304         NAPI_GRO_CB(skb)->flush = 1;
 305         return NULL;
 306 }
 307
 308 int udp_gro_complete(struct sk_buff *skb, int nhoff)
 309 {
 310         struct udp_offload_priv *uo_priv;
 311         __be16 newlen = htons(skb->len - nhoff);
 312         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 313         int err = -ENOSYS;
 314
 315         uh->len = newlen;
 316
 317         rcu_read_lock();
 318
 319         uo_priv = rcu_dereference(udp_offload_base);
 320         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 321                 if (uo_priv->offload->port == uh->dest &&
 322                     uo_priv->offload->callbacks.gro_complete)
 323                         break;
 324         }
 325
 326         if (uo_priv != NULL) {
 327                 NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
 328                 err = uo_priv->offload->callbacks.gro_complete(skb, nhoff + sizeof(struct udphdr));
 329         }
 330
 331         rcu_read_unlock();
 332         return err;
 333 }
 334
 335 static int udp4_gro_complete(struct sk_buff *skb, int nhoff)
 336 {
 337         const struct iphdr *iph = ip_hdr(skb);
 338         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 339
 340         if (uh->check)
 341                 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
 342                                           iph->daddr, 0);
 343
 344         return udp_gro_complete(skb, nhoff);
 345 }
 346
 347 static const struct net_offload udpv4_offload = {
 348         .callbacks = {
 349                 .gso_send_check = udp4_ufo_send_check,
 350                 .gso_segment = udp4_ufo_fragment,
 351                 .gro_receive  = udp4_gro_receive,
 352                 .gro_complete = udp4_gro_complete,
 353         },
 354 };
 355
 356 int __init udpv4_offload_init(void)
 357 {
 358         return inet_add_offload(&udpv4_offload, IPPROTO_UDP);
 359 }