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 static void tcp_gso_tstamp(struct sk_buff *skb, unsigned int ts_seq,
  18                            unsigned int seq, unsigned int mss)
  19 {
  20         while (skb) {
  21                 if (before(ts_seq, seq + mss)) {
  22                         skb_shinfo(skb)->tx_flags |= SKBTX_SW_TSTAMP;
  23                         skb_shinfo(skb)->tskey = ts_seq;
  24                         return;
  25                 }
  26
  27                 skb = skb->next;
  28                 seq += mss;
  29         }
  30 }
  31
  32 struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
  33                                 netdev_features_t features)
  34 {
  35         struct sk_buff *segs = ERR_PTR(-EINVAL);
  36         unsigned int sum_truesize = 0;
  37         struct tcphdr *th;
  38         unsigned int thlen;
  39         unsigned int seq;
  40         __be32 delta;
  41         unsigned int oldlen;
  42         unsigned int mss;
  43         struct sk_buff *gso_skb = skb;
  44         __sum16 newcheck;
  45         bool ooo_okay, copy_destructor;
  46
  47         if (!pskb_may_pull(skb, sizeof(*th)))
  48                 goto out;
  49
  50         th = tcp_hdr(skb);
  51         thlen = th->doff * 4;
  52         if (thlen < sizeof(*th))
  53                 goto out;
  54
  55         if (!pskb_may_pull(skb, thlen))
  56                 goto out;
  57
  58         oldlen = (u16)~skb->len;
  59         __skb_pull(skb, thlen);
  60
  61         mss = tcp_skb_mss(skb);
  62         if (unlikely(skb->len <= mss))
  63                 goto out;
  64
  65         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
  66                 /* Packet is from an untrusted source, reset gso_segs. */
  67                 int type = skb_shinfo(skb)->gso_type;
  68
  69                 if (unlikely(type &
  70                              ~(SKB_GSO_TCPV4 |
  71                                SKB_GSO_DODGY |
  72                                SKB_GSO_TCP_ECN |
  73                                SKB_GSO_TCPV6 |
  74                                SKB_GSO_GRE |
  75                                SKB_GSO_GRE_CSUM |
  76                                SKB_GSO_IPIP |
  77                                SKB_GSO_SIT |
  78                                SKB_GSO_MPLS |
  79                                SKB_GSO_UDP_TUNNEL |
  80                                SKB_GSO_UDP_TUNNEL_CSUM |
  81                                0) ||
  82                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
  83                         goto out;
  84
  85                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
  86
  87                 segs = NULL;
  88                 goto out;
  89         }
  90
  91         copy_destructor = gso_skb->destructor == tcp_wfree;
  92         ooo_okay = gso_skb->ooo_okay;
  93         /* All segments but the first should have ooo_okay cleared */
  94         skb->ooo_okay = 0;
  95
  96         segs = skb_segment(skb, features);
  97         if (IS_ERR(segs))
  98                 goto out;
  99
 100         /* Only first segment might have ooo_okay set */
 101         segs->ooo_okay = ooo_okay;
 102
 103         delta = htonl(oldlen + (thlen + mss));
 104
 105         skb = segs;
 106         th = tcp_hdr(skb);
 107         seq = ntohl(th->seq);
 108
 109         if (unlikely(skb_shinfo(gso_skb)->tx_flags & SKBTX_SW_TSTAMP))
 110                 tcp_gso_tstamp(segs, skb_shinfo(gso_skb)->tskey, seq, mss);
 111
 112         newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
 113                                                (__force u32)delta));
 114
 115         do {
 116                 th->fin = th->psh = 0;
 117                 th->check = newcheck;
 118
 119                 if (skb->ip_summed != CHECKSUM_PARTIAL)
 120                         th->check = gso_make_checksum(skb, ~th->check);
 121
 122                 seq += mss;
 123                 if (copy_destructor) {
 124                         skb->destructor = gso_skb->destructor;
 125                         skb->sk = gso_skb->sk;
 126                         sum_truesize += skb->truesize;
 127                 }
 128                 skb = skb->next;
 129                 th = tcp_hdr(skb);
 130
 131                 th->seq = htonl(seq);
 132                 th->cwr = 0;
 133         } while (skb->next);
 134
 135         /* Following permits TCP Small Queues to work well with GSO :
 136          * The callback to TCP stack will be called at the time last frag
 137          * is freed at TX completion, and not right now when gso_skb
 138          * is freed by GSO engine
 139          */
 140         if (copy_destructor) {
 141                 swap(gso_skb->sk, skb->sk);
 142                 swap(gso_skb->destructor, skb->destructor);
 143                 sum_truesize += skb->truesize;
 144                 atomic_add(sum_truesize - gso_skb->truesize,
 145                            &skb->sk->sk_wmem_alloc);
 146         }
 147
 148         delta = htonl(oldlen + (skb_tail_pointer(skb) -
 149                                 skb_transport_header(skb)) +
 150                       skb->data_len);
 151         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
 152                                 (__force u32)delta));
 153         if (skb->ip_summed != CHECKSUM_PARTIAL)
 154                 th->check = gso_make_checksum(skb, ~th->check);
 155 out:
 156         return segs;
 157 }
 158
 159 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
 160 {
 161         struct sk_buff **pp = NULL;
 162         struct sk_buff *p;
 163         struct tcphdr *th;
 164         struct tcphdr *th2;
 165         unsigned int len;
 166         unsigned int thlen;
 167         __be32 flags;
 168         unsigned int mss = 1;
 169         unsigned int hlen;
 170         unsigned int off;
 171         int flush = 1;
 172         int i;
 173
 174         off = skb_gro_offset(skb);
 175         hlen = off + sizeof(*th);
 176         th = skb_gro_header_fast(skb, off);
 177         if (skb_gro_header_hard(skb, hlen)) {
 178                 th = skb_gro_header_slow(skb, hlen, off);
 179                 if (unlikely(!th))
 180                         goto out;
 181         }
 182
 183         thlen = th->doff * 4;
 184         if (thlen < sizeof(*th))
 185                 goto out;
 186
 187         hlen = off + thlen;
 188         if (skb_gro_header_hard(skb, hlen)) {
 189                 th = skb_gro_header_slow(skb, hlen, off);
 190                 if (unlikely(!th))
 191                         goto out;
 192         }
 193
 194         skb_gro_pull(skb, thlen);
 195
 196         len = skb_gro_len(skb);
 197         flags = tcp_flag_word(th);
 198
 199         for (; (p = *head); head = &p->next) {
 200                 if (!NAPI_GRO_CB(p)->same_flow)
 201                         continue;
 202
 203                 th2 = tcp_hdr(p);
 204
 205                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
 206                         NAPI_GRO_CB(p)->same_flow = 0;
 207                         continue;
 208                 }
 209
 210                 goto found;
 211         }
 212
 213         goto out_check_final;
 214
 215 found:
 216         /* Include the IP ID check below from the inner most IP hdr */
 217         flush = NAPI_GRO_CB(p)->flush | NAPI_GRO_CB(p)->flush_id;
 218         flush |= (__force int)(flags & TCP_FLAG_CWR);
 219         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
 220                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
 221         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
 222         for (i = sizeof(*th); i < thlen; i += 4)
 223                 flush |= *(u32 *)((u8 *)th + i) ^
 224                          *(u32 *)((u8 *)th2 + i);
 225
 226         mss = tcp_skb_mss(p);
 227
 228         flush |= (len - 1) >= mss;
 229         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
 230
 231         if (flush || skb_gro_receive(head, skb)) {
 232                 mss = 1;
 233                 goto out_check_final;
 234         }
 235
 236         p = *head;
 237         th2 = tcp_hdr(p);
 238         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
 239
 240 out_check_final:
 241         flush = len < mss;
 242         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
 243                                         TCP_FLAG_RST | TCP_FLAG_SYN |
 244                                         TCP_FLAG_FIN));
 245
 246         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
 247                 pp = head;
 248
 249 out:
 250         NAPI_GRO_CB(skb)->flush |= (flush != 0);
 251
 252         return pp;
 253 }
 254
 255 int tcp_gro_complete(struct sk_buff *skb)
 256 {
 257         struct tcphdr *th = tcp_hdr(skb);
 258
 259         skb->csum_start = (unsigned char *)th - skb->head;
 260         skb->csum_offset = offsetof(struct tcphdr, check);
 261         skb->ip_summed = CHECKSUM_PARTIAL;
 262
 263         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
 264
 265         if (th->cwr)
 266                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 267
 268         return 0;
 269 }
 270 EXPORT_SYMBOL(tcp_gro_complete);
 271
 272 static int tcp_v4_gso_send_check(struct sk_buff *skb)
 273 {
 274         const struct iphdr *iph;
 275         struct tcphdr *th;
 276
 277         if (!pskb_may_pull(skb, sizeof(*th)))
 278                 return -EINVAL;
 279
 280         iph = ip_hdr(skb);
 281         th = tcp_hdr(skb);
 282
 283         th->check = 0;
 284         skb->ip_summed = CHECKSUM_PARTIAL;
 285         __tcp_v4_send_check(skb, iph->saddr, iph->daddr);
 286         return 0;
 287 }
 288
 289 static struct sk_buff **tcp4_gro_receive(struct sk_buff **head, struct sk_buff *skb)
 290 {
 291         /* Use the IP hdr immediately proceeding for this transport */
 292         const struct iphdr *iph = skb_gro_network_header(skb);
 293         __wsum wsum;
 294
 295         /* Don't bother verifying checksum if we're going to flush anyway. */
 296         if (NAPI_GRO_CB(skb)->flush)
 297                 goto skip_csum;
 298
 299         wsum = NAPI_GRO_CB(skb)->csum;
 300
 301         switch (skb->ip_summed) {
 302         case CHECKSUM_NONE:
 303                 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb),
 304                                     0);
 305
 306                 /* fall through */
 307
 308         case CHECKSUM_COMPLETE:
 309                 if (!tcp_v4_check(skb_gro_len(skb), iph->saddr, iph->daddr,
 310                                   wsum)) {
 311                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 312                         break;
 313                 }
 314
 315                 NAPI_GRO_CB(skb)->flush = 1;
 316                 return NULL;
 317         }
 318
 319 skip_csum:
 320         return tcp_gro_receive(head, skb);
 321 }
 322
 323 static int tcp4_gro_complete(struct sk_buff *skb, int thoff)
 324 {
 325         const struct iphdr *iph = ip_hdr(skb);
 326         struct tcphdr *th = tcp_hdr(skb);
 327
 328         th->check = ~tcp_v4_check(skb->len - thoff, iph->saddr,
 329                                   iph->daddr, 0);
 330         skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
 331
 332         return tcp_gro_complete(skb);
 333 }
 334
 335 static const struct net_offload tcpv4_offload = {
 336         .callbacks = {
 337                 .gso_send_check =       tcp_v4_gso_send_check,
 338                 .gso_segment    =       tcp_gso_segment,
 339                 .gro_receive    =       tcp4_gro_receive,
 340                 .gro_complete   =       tcp4_gro_complete,
 341         },
 342 };
 343
 344 int __init tcpv4_offload_init(void)
 345 {
 346         return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);
 347 }