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