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         if (!pskb_may_pull(skb, sizeof(struct udphdr)))
  31                 return -EINVAL;
  32
  33         if (likely(!skb->encapsulation)) {
  34                 const struct iphdr *iph;
  35                 struct udphdr *uh;
  36
  37                 iph = ip_hdr(skb);
  38                 uh = udp_hdr(skb);
  39
  40                 uh->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr, skb->len,
  41                                 IPPROTO_UDP, 0);
  42                 skb->csum_start = skb_transport_header(skb) - skb->head;
  43                 skb->csum_offset = offsetof(struct udphdr, check);
  44                 skb->ip_summed = CHECKSUM_PARTIAL;
  45         }
  46
  47         return 0;
  48 }
  49
  50 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
  51                                        netdev_features_t features)
  52 {
  53         struct sk_buff *segs = ERR_PTR(-EINVAL);
  54         u16 mac_offset = skb->mac_header;
  55         int mac_len = skb->mac_len;
  56         int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
  57         __be16 protocol = skb->protocol;
  58         netdev_features_t enc_features;
  59         int udp_offset, outer_hlen;
  60         unsigned int oldlen;
  61         bool need_csum;
  62
  63         oldlen = (u16)~skb->len;
  64
  65         if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
  66                 goto out;
  67
  68         skb->encapsulation = 0;
  69         __skb_pull(skb, tnl_hlen);
  70         skb_reset_mac_header(skb);
  71         skb_set_network_header(skb, skb_inner_network_offset(skb));
  72         skb->mac_len = skb_inner_network_offset(skb);
  73         skb->protocol = htons(ETH_P_TEB);
  74
  75         need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
  76         if (need_csum)
  77                 skb->encap_hdr_csum = 1;
  78
  79         /* segment inner packet. */
  80         enc_features = skb->dev->hw_enc_features & netif_skb_features(skb);
  81         segs = skb_mac_gso_segment(skb, enc_features);
  82         if (IS_ERR_OR_NULL(segs)) {
  83                 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
  84                                      mac_len);
  85                 goto out;
  86         }
  87
  88         outer_hlen = skb_tnl_header_len(skb);
  89         udp_offset = outer_hlen - tnl_hlen;
  90         skb = segs;
  91         do {
  92                 struct udphdr *uh;
  93                 int len;
  94
  95                 skb_reset_inner_headers(skb);
  96                 skb->encapsulation = 1;
  97
  98                 skb->mac_len = mac_len;
  99
 100                 skb_push(skb, outer_hlen);
 101                 skb_reset_mac_header(skb);
 102                 skb_set_network_header(skb, mac_len);
 103                 skb_set_transport_header(skb, udp_offset);
 104                 len = skb->len - udp_offset;
 105                 uh = udp_hdr(skb);
 106                 uh->len = htons(len);
 107
 108                 if (need_csum) {
 109                         __be32 delta = htonl(oldlen + len);
 110
 111                         uh->check = ~csum_fold((__force __wsum)
 112                                                ((__force u32)uh->check +
 113                                                 (__force u32)delta));
 114                         uh->check = gso_make_checksum(skb, ~uh->check);
 115
 116                         if (uh->check == 0)
 117                                 uh->check = CSUM_MANGLED_0;
 118                 }
 119
 120                 skb->protocol = protocol;
 121         } while ((skb = skb->next));
 122 out:
 123         return segs;
 124 }
 125
 126 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
 127                                          netdev_features_t features)
 128 {
 129         struct sk_buff *segs = ERR_PTR(-EINVAL);
 130         unsigned int mss;
 131         int offset;
 132         __wsum csum;
 133
 134         if (skb->encapsulation &&
 135             (skb_shinfo(skb)->gso_type &
 136              (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
 137                 segs = skb_udp_tunnel_segment(skb, features);
 138                 goto out;
 139         }
 140
 141         mss = skb_shinfo(skb)->gso_size;
 142         if (unlikely(skb->len <= mss))
 143                 goto out;
 144
 145         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
 146                 /* Packet is from an untrusted source, reset gso_segs. */
 147                 int type = skb_shinfo(skb)->gso_type;
 148
 149                 if (unlikely(type & ~(SKB_GSO_UDP | SKB_GSO_DODGY |
 150                                       SKB_GSO_UDP_TUNNEL |
 151                                       SKB_GSO_UDP_TUNNEL_CSUM |
 152                                       SKB_GSO_IPIP |
 153                                       SKB_GSO_GRE | SKB_GSO_GRE_CSUM |
 154                                       SKB_GSO_MPLS) ||
 155                              !(type & (SKB_GSO_UDP))))
 156                         goto out;
 157
 158                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
 159
 160                 segs = NULL;
 161                 goto out;
 162         }
 163
 164         /* Do software UFO. Complete and fill in the UDP checksum as
 165          * HW cannot do checksum of UDP packets sent as multiple
 166          * IP fragments.
 167          */
 168         offset = skb_checksum_start_offset(skb);
 169         csum = skb_checksum(skb, offset, skb->len - offset, 0);
 170         offset += skb->csum_offset;
 171         *(__sum16 *)(skb->data + offset) = csum_fold(csum);
 172         skb->ip_summed = CHECKSUM_NONE;
 173
 174         /* Fragment the skb. IP headers of the fragments are updated in
 175          * inet_gso_segment()
 176          */
 177         segs = skb_segment(skb, features);
 178 out:
 179         return segs;
 180 }
 181
 182 int udp_add_offload(struct udp_offload *uo)
 183 {
 184         struct udp_offload_priv *new_offload = kzalloc(sizeof(*new_offload), GFP_ATOMIC);
 185
 186         if (!new_offload)
 187                 return -ENOMEM;
 188
 189         new_offload->offload = uo;
 190
 191         spin_lock(&udp_offload_lock);
 192         new_offload->next = udp_offload_base;
 193         rcu_assign_pointer(udp_offload_base, new_offload);
 194         spin_unlock(&udp_offload_lock);
 195
 196         return 0;
 197 }
 198 EXPORT_SYMBOL(udp_add_offload);
 199
 200 static void udp_offload_free_routine(struct rcu_head *head)
 201 {
 202         struct udp_offload_priv *ou_priv = container_of(head, struct udp_offload_priv, rcu);
 203         kfree(ou_priv);
 204 }
 205
 206 void udp_del_offload(struct udp_offload *uo)
 207 {
 208         struct udp_offload_priv __rcu **head = &udp_offload_base;
 209         struct udp_offload_priv *uo_priv;
 210
 211         spin_lock(&udp_offload_lock);
 212
 213         uo_priv = udp_deref_protected(*head);
 214         for (; uo_priv != NULL;
 215              uo_priv = udp_deref_protected(*head)) {
 216                 if (uo_priv->offload == uo) {
 217                         rcu_assign_pointer(*head,
 218                                            udp_deref_protected(uo_priv->next));
 219                         goto unlock;
 220                 }
 221                 head = &uo_priv->next;
 222         }
 223         pr_warn("udp_del_offload: didn't find offload for port %d\n", ntohs(uo->port));
 224 unlock:
 225         spin_unlock(&udp_offload_lock);
 226         if (uo_priv != NULL)
 227                 call_rcu(&uo_priv->rcu, udp_offload_free_routine);
 228 }
 229 EXPORT_SYMBOL(udp_del_offload);
 230
 231 static struct sk_buff **udp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
 232 {
 233         struct udp_offload_priv *uo_priv;
 234         struct sk_buff *p, **pp = NULL;
 235         struct udphdr *uh, *uh2;
 236         unsigned int hlen, off;
 237         int flush = 1;
 238
 239         if (NAPI_GRO_CB(skb)->udp_mark ||
 240             (!skb->encapsulation && skb->ip_summed != CHECKSUM_COMPLETE))
 241                 goto out;
 242
 243         /* mark that this skb passed once through the udp gro layer */
 244         NAPI_GRO_CB(skb)->udp_mark = 1;
 245
 246         off  = skb_gro_offset(skb);
 247         hlen = off + sizeof(*uh);
 248         uh   = skb_gro_header_fast(skb, off);
 249         if (skb_gro_header_hard(skb, hlen)) {
 250                 uh = skb_gro_header_slow(skb, hlen, off);
 251                 if (unlikely(!uh))
 252                         goto out;
 253         }
 254
 255         rcu_read_lock();
 256         uo_priv = rcu_dereference(udp_offload_base);
 257         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 258                 if (uo_priv->offload->port == uh->dest &&
 259                     uo_priv->offload->callbacks.gro_receive)
 260                         goto unflush;
 261         }
 262         goto out_unlock;
 263
 264 unflush:
 265         flush = 0;
 266
 267         for (p = *head; p; p = p->next) {
 268                 if (!NAPI_GRO_CB(p)->same_flow)
 269                         continue;
 270
 271                 uh2 = (struct udphdr   *)(p->data + off);
 272                 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) {
 273                         NAPI_GRO_CB(p)->same_flow = 0;
 274                         continue;
 275                 }
 276         }
 277
 278         skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
 279         skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
 280         pp = uo_priv->offload->callbacks.gro_receive(head, skb);
 281
 282 out_unlock:
 283         rcu_read_unlock();
 284 out:
 285         NAPI_GRO_CB(skb)->flush |= flush;
 286         return pp;
 287 }
 288
 289 static int udp_gro_complete(struct sk_buff *skb, int nhoff)
 290 {
 291         struct udp_offload_priv *uo_priv;
 292         __be16 newlen = htons(skb->len - nhoff);
 293         struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
 294         int err = -ENOSYS;
 295
 296         uh->len = newlen;
 297
 298         rcu_read_lock();
 299
 300         uo_priv = rcu_dereference(udp_offload_base);
 301         for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
 302                 if (uo_priv->offload->port == uh->dest &&
 303                     uo_priv->offload->callbacks.gro_complete)
 304                         break;
 305         }
 306
 307         if (uo_priv != NULL)
 308                 err = uo_priv->offload->callbacks.gro_complete(skb, nhoff + sizeof(struct udphdr));
 309
 310         rcu_read_unlock();
 311         return err;
 312 }
 313
 314 static const struct net_offload udpv4_offload = {
 315         .callbacks = {
 316                 .gso_send_check = udp4_ufo_send_check,
 317                 .gso_segment = udp4_ufo_fragment,
 318                 .gro_receive  = udp_gro_receive,
 319                 .gro_complete = udp_gro_complete,
 320         },
 321 };
 322
 323 int __init udpv4_offload_init(void)
 324 {
 325         return inet_add_offload(&udpv4_offload, IPPROTO_UDP);
 326 }