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