net/ipv4/tcp_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  *      TCPv4 GSO/GRO support
  11  */
  12
  13 #include <linux/skbuff.h>
  14 #include <net/tcp.h>
  15 #include <net/protocol.h>
  16
  17 struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
  18                                 netdev_features_t features)
  19 {
  20         struct sk_buff *segs = ERR_PTR(-EINVAL);
  21         unsigned int sum_truesize = 0;
  22         struct tcphdr *th;
  23         unsigned int thlen;
  24         unsigned int seq;
  25         __be32 delta;
  26         unsigned int oldlen;
  27         unsigned int mss;
  28         struct sk_buff *gso_skb = skb;
  29         __sum16 newcheck;
  30         bool ooo_okay, copy_destructor;
  31
  32         if (!pskb_may_pull(skb, sizeof(*th)))
  33                 goto out;
  34
  35         th = tcp_hdr(skb);
  36         thlen = th->doff * 4;
  37         if (thlen < sizeof(*th))
  38                 goto out;
  39
  40         if (!pskb_may_pull(skb, thlen))
  41                 goto out;
  42
  43         oldlen = (u16)~skb->len;
  44         __skb_pull(skb, thlen);
  45
  46         mss = tcp_skb_mss(skb);
  47         if (unlikely(skb->len <= mss))
  48                 goto out;
  49
  50         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
  51                 /* Packet is from an untrusted source, reset gso_segs. */
  52                 int type = skb_shinfo(skb)->gso_type;
  53
  54                 if (unlikely(type &
  55                              ~(SKB_GSO_TCPV4 |
  56                                SKB_GSO_DODGY |
  57                                SKB_GSO_TCP_ECN |
  58                                SKB_GSO_TCPV6 |
  59                                SKB_GSO_GRE |
  60                                SKB_GSO_IPIP |
  61                                SKB_GSO_SIT |
  62                                SKB_GSO_MPLS |
  63                                SKB_GSO_UDP_TUNNEL |
  64                                SKB_GSO_UDP_TUNNEL_CSUM |
  65                                0) ||
  66                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
  67                         goto out;
  68
  69                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
  70
  71                 segs = NULL;
  72                 goto out;
  73         }
  74
  75         copy_destructor = gso_skb->destructor == tcp_wfree;
  76         ooo_okay = gso_skb->ooo_okay;
  77         /* All segments but the first should have ooo_okay cleared */
  78         skb->ooo_okay = 0;
  79
  80         segs = skb_segment(skb, features);
  81         if (IS_ERR(segs))
  82                 goto out;
  83
  84         /* Only first segment might have ooo_okay set */
  85         segs->ooo_okay = ooo_okay;
  86
  87         delta = htonl(oldlen + (thlen + mss));
  88
  89         skb = segs;
  90         th = tcp_hdr(skb);
  91         seq = ntohl(th->seq);
  92
  93         newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
  94                                                (__force u32)delta));
  95
  96         do {
  97                 th->fin = th->psh = 0;
  98                 th->check = newcheck;
  99
 100                 if (skb->ip_summed != CHECKSUM_PARTIAL)
 101                         th->check = gso_make_checksum(skb, ~th->check);
 102
 103                 seq += mss;
 104                 if (copy_destructor) {
 105                         skb->destructor = gso_skb->destructor;
 106                         skb->sk = gso_skb->sk;
 107                         sum_truesize += skb->truesize;
 108                 }
 109                 skb = skb->next;
 110                 th = tcp_hdr(skb);
 111
 112                 th->seq = htonl(seq);
 113                 th->cwr = 0;
 114         } while (skb->next);
 115
 116         /* Following permits TCP Small Queues to work well with GSO :
 117          * The callback to TCP stack will be called at the time last frag
 118          * is freed at TX completion, and not right now when gso_skb
 119          * is freed by GSO engine
 120          */
 121         if (copy_destructor) {
 122                 swap(gso_skb->sk, skb->sk);
 123                 swap(gso_skb->destructor, skb->destructor);
 124                 sum_truesize += skb->truesize;
 125                 atomic_add(sum_truesize - gso_skb->truesize,
 126                            &skb->sk->sk_wmem_alloc);
 127         }
 128
 129         delta = htonl(oldlen + (skb_tail_pointer(skb) -
 130                                 skb_transport_header(skb)) +
 131                       skb->data_len);
 132         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
 133                                 (__force u32)delta));
 134         if (skb->ip_summed != CHECKSUM_PARTIAL)
 135                 th->check = gso_make_checksum(skb, ~th->check);
 136 out:
 137         return segs;
 138 }
 139
 140 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
 141 {
 142         struct sk_buff **pp = NULL;
 143         struct sk_buff *p;
 144         struct tcphdr *th;
 145         struct tcphdr *th2;
 146         unsigned int len;
 147         unsigned int thlen;
 148         __be32 flags;
 149         unsigned int mss = 1;
 150         unsigned int hlen;
 151         unsigned int off;
 152         int flush = 1;
 153         int i;
 154
 155         off = skb_gro_offset(skb);
 156         hlen = off + sizeof(*th);
 157         th = skb_gro_header_fast(skb, off);
 158         if (skb_gro_header_hard(skb, hlen)) {
 159                 th = skb_gro_header_slow(skb, hlen, off);
 160                 if (unlikely(!th))
 161                         goto out;
 162         }
 163
 164         thlen = th->doff * 4;
 165         if (thlen < sizeof(*th))
 166                 goto out;
 167
 168         hlen = off + thlen;
 169         if (skb_gro_header_hard(skb, hlen)) {
 170                 th = skb_gro_header_slow(skb, hlen, off);
 171                 if (unlikely(!th))
 172                         goto out;
 173         }
 174
 175         skb_gro_pull(skb, thlen);
 176
 177         len = skb_gro_len(skb);
 178         flags = tcp_flag_word(th);
 179
 180         for (; (p = *head); head = &p->next) {
 181                 if (!NAPI_GRO_CB(p)->same_flow)
 182                         continue;
 183
 184                 th2 = tcp_hdr(p);
 185
 186                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
 187                         NAPI_GRO_CB(p)->same_flow = 0;
 188                         continue;
 189                 }
 190
 191                 goto found;
 192         }
 193
 194         goto out_check_final;
 195
 196 found:
 197         /* Include the IP ID check below from the inner most IP hdr */
 198         flush = NAPI_GRO_CB(p)->flush | NAPI_GRO_CB(p)->flush_id;
 199         flush |= (__force int)(flags & TCP_FLAG_CWR);
 200         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
 201                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
 202         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
 203         for (i = sizeof(*th); i < thlen; i += 4)
 204                 flush |= *(u32 *)((u8 *)th + i) ^
 205                          *(u32 *)((u8 *)th2 + i);
 206
 207         mss = tcp_skb_mss(p);
 208
 209         flush |= (len - 1) >= mss;
 210         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
 211
 212         if (flush || skb_gro_receive(head, skb)) {
 213                 mss = 1;
 214                 goto out_check_final;
 215         }
 216
 217         p = *head;
 218         th2 = tcp_hdr(p);
 219         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
 220
 221 out_check_final:
 222         flush = len < mss;
 223         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
 224                                         TCP_FLAG_RST | TCP_FLAG_SYN |
 225                                         TCP_FLAG_FIN));
 226
 227         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
 228                 pp = head;
 229
 230 out:
 231         NAPI_GRO_CB(skb)->flush |= (flush != 0);
 232
 233         return pp;
 234 }
 235
 236 int tcp_gro_complete(struct sk_buff *skb)
 237 {
 238         struct tcphdr *th = tcp_hdr(skb);
 239
 240         skb->csum_start = (unsigned char *)th - skb->head;
 241         skb->csum_offset = offsetof(struct tcphdr, check);
 242         skb->ip_summed = CHECKSUM_PARTIAL;
 243
 244         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
 245
 246         if (th->cwr)
 247                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 248
 249         return 0;
 250 }
 251 EXPORT_SYMBOL(tcp_gro_complete);
 252
 253 static int tcp_v4_gso_send_check(struct sk_buff *skb)
 254 {
 255         const struct iphdr *iph;
 256         struct tcphdr *th;
 257
 258         if (!pskb_may_pull(skb, sizeof(*th)))
 259                 return -EINVAL;
 260
 261         iph = ip_hdr(skb);
 262         th = tcp_hdr(skb);
 263
 264         th->check = 0;
 265         skb->ip_summed = CHECKSUM_PARTIAL;
 266         __tcp_v4_send_check(skb, iph->saddr, iph->daddr);
 267         return 0;
 268 }
 269
 270 static struct sk_buff **tcp4_gro_receive(struct sk_buff **head, struct sk_buff *skb)
 271 {
 272         /* Use the IP hdr immediately proceeding for this transport */
 273         const struct iphdr *iph = skb_gro_network_header(skb);
 274         __wsum wsum;
 275
 276         /* Don't bother verifying checksum if we're going to flush anyway. */
 277         if (NAPI_GRO_CB(skb)->flush)
 278                 goto skip_csum;
 279
 280         wsum = NAPI_GRO_CB(skb)->csum;
 281
 282         switch (skb->ip_summed) {
 283         case CHECKSUM_NONE:
 284                 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb),
 285                                     0);
 286
 287                 /* fall through */
 288
 289         case CHECKSUM_COMPLETE:
 290                 if (!tcp_v4_check(skb_gro_len(skb), iph->saddr, iph->daddr,
 291                                   wsum)) {
 292                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 293                         break;
 294                 }
 295
 296                 NAPI_GRO_CB(skb)->flush = 1;
 297                 return NULL;
 298         }
 299
 300 skip_csum:
 301         return tcp_gro_receive(head, skb);
 302 }
 303
 304 static int tcp4_gro_complete(struct sk_buff *skb, int thoff)
 305 {
 306         const struct iphdr *iph = ip_hdr(skb);
 307         struct tcphdr *th = tcp_hdr(skb);
 308
 309         th->check = ~tcp_v4_check(skb->len - thoff, iph->saddr,
 310                                   iph->daddr, 0);
 311         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 312
 313         return tcp_gro_complete(skb);
 314 }
 315
 316 static const struct net_offload tcpv4_offload = {
 317         .callbacks = {
 318                 .gso_send_check =       tcp_v4_gso_send_check,
 319                 .gso_segment    =       tcp_gso_segment,
 320                 .gro_receive    =       tcp4_gro_receive,
 321                 .gro_complete   =       tcp4_gro_complete,
 322         },
 323 };
 324
 325 int __init tcpv4_offload_init(void)
 326 {
 327         return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);
 328 }